将 AI 代理注入 CI/CD：在 GitHub Actions 中使用 GitHub Copilot CLI 实现智能故障

Source: Dev.to

TL;DR for the Busy Dev

我们已经习惯了在语法错误或单元测试失败时 CI/CD 流水线会中止。通过将 GitHub Copilot CLI 直接嵌入 GitHub Action，你可以创建 AI 代理来审查代码的安全性、逻辑或产品规范合规性。如果代理检测到关键问题，它会以编程方式使工作流失败，在人工审查之前就阻止合并。

DevOps 的“左移”(Shift‑Left) 目标——尽可能早地捕获问题——已经在确定性问题（代码检查器、测试运行器）上得到实现。非确定性审查（SQL 注入安全、文档更新、验收标准合规）仍然依赖人工。本文展示了如何使用 安全代理 在 CI 流水线中运行、利用 Copilot CLI 扫描代码，并在出现关键漏洞时使构建失败，从而弥合这一差距。

Building a Security Agent in CI

此工作流中最关键的部分不是 YAML，而是提示工程。AI 必须充当严苛的审计员，而不是乐于助人的助手。提示存放在 .github/agents/security-reporter.agent.md 中。
Link to full prompt file here。

Prompt File (security-reporter.agent.md)

name: SecurityReportAgent
description: Security Report Agent - Analyzes TypeScript and React code for security vulnerabilities and creates security reports
model: GPT-5.1 (Preview)

## Purpose
This agent performs comprehensive security analysis of the Astro, TypeScript code. It identifies security vulnerabilities, assesses risks, and produces detailed security reports without modifying the codebase directly.

## Security Scanning Capabilities

### Code Analysis
- **SAST (Static Code Analysis)** – scans TypeScript/React source for:
  - SQL Injection, XSS, CSRF
  - Authentication/authorization flaws
  - Insecure cryptographic implementations
  - Hard‑coded secrets, path traversal, insecure deserialization
  - Input validation, data encryption, error handling, missing security headers
  - Dependency vulnerabilities, information disclosure risks

### Dependency & Component Analysis
- **SCA (Software Composition Analysis)** – monitors npm dependencies for known CVEs
- License scanning, outdated software detection, malware detection in the supply chain

### Infrastructure & Configuration
- Secrets detection, cloud configuration review (Azure Functions), IaC scanning (Terraform/CloudFormation/K8s), container image scanning

### API & Runtime Security
- API security, database security, WebSocket security, file‑upload security

### Compliance & Best Practices
- OWASP Top 10, TypeScript/React security guidelines, secure coding standards, security‑header verification, GDPR/privacy considerations

### Security Metrics & Reporting
- Vulnerability count by severity, code‑coverage analysis, OWASP mapping, CWE classification, risk score, remediation timeline

Report Structure

Security Assessment Report

1. Executive Summary

   • Security Posture: [Risk Level] (e.g., HIGH RISK)
   • Score: [0‑10]/10
   • Findings Summary

   Severity	Count
   Critical	[Count]
   High	[Count]
   Medium	[Count]
   Low	[Count]

2. Vulnerability Findings (repeat for each issue)

   • Severity: Critical/High/Medium/Low
   • Category: (e.g., Injection, Authentication)
   • Location: file and line number
   • Description, Impact, Recommendation, References

3. Security Best Practices Review – what follows best practices, what needs improvement, configuration recommendations.

4. Dependency Analysis – vulnerable packages and suggested updates.

5. Action Items – prioritized fix list (quick wins vs. complex remediation).

6. Intentional Vulnerabilities – list any critical/high findings in:

   • Any file under infra/
   • Any path containing legacy-vibe
     Mark them as “Intentional – No Action Required.”

7. Critical Vulnerability Warning

   • Review all CRITICAL findings.

   • Exclude any that appear in the “Intentional Vulnerabilities” paths above.

   • If any remaining critical vulnerabilities exist:

     ### Blocking Critical Vulnerabilities
     *[brief list of remaining critical issues]*
     
     THIS ASSESSMENT CONTAINS A CRITICAL VULNERABILITY

   • Do not modify the warning message.

   • If all critical findings are filtered out, omit the warning entirely.

The pipeline treats the exact string THIS ASSESSMENT CONTAINS A CRITICAL VULNERABILITY as a kill switch. Its presence causes the workflow to fail; its absence allows the run to continue.

GitHub Actions Implementation

Below is a minimal workflow that installs the Copilot CLI, runs the security agent, and fails on the kill‑switch string.

name: Security Scan

on:
  pull_request:
    branches: [ main ]

jobs:
  security-report:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      actions: read
      security-events: write

    steps:
      - name: Checkout repository
        uses: actions/checkout@v4

      - name: Set up Node
        uses: actions/setup-node@v4
        with:
          node-version: '20'

      - name: Install Copilot CLI
        run: npm i -g @github/copilot

      - name: Run Security Agent
        env:
          COPILOT_TOKEN: ${{ secrets.COPILOT_PAT }}   # fine‑grained PAT with “Copilot Requests: Read”
        run: |
          copilot run \
            --prompt-file .github/agents/security-reporter.agent.md \
            --repo ${{ github.repository }} \
            > agent-output.txt

      - name: Check for critical warning
        id: check
        run: |
          if grep -q "THIS ASSESSMENT CONTAINS A CRITICAL VULNERABILITY" agent-output.txt; then
            echo "critical=true" >> $GITHUB_OUTPUT
          else
            echo "critical=false" >> $GITHUB_OUTPUT
          fi

      - name: Fail on critical vulnerabilities
        if: steps.check.outputs.critical == 'true'
        run: |
          echo "🚨 Critical security issues detected – failing the workflow."
          exit 1

      - name: Upload report as artifact
        if: steps.check.outputs.critical == 'false'
        uses: actions/upload-artifact@v4
        with:
          name: security-report
          path: agent-output.txt

Key Points

• PAT requirement – Create a fine‑grained Personal Access Token with Copilot Requests: Read permission and store it as COPILOT_PAT in repository secrets.
• The Bash step captures the agent’s output, searches for the exact kill‑switch string, and sets an output flag.
• A subsequent step fails the job (exit 1) when the flag is true.
• When no critical issues remain, the report is uploaded as an artifact for review.

With this setup, AI‑driven security reviews become an integral, automated part of your CI/CD pipeline, turning natural‑language analysis into a deterministic pass/fail signal. 🚀