AI Agent 当贡献者：KubeStellar 81% PR 接受率背后的工程实践

# .github/workflows/ai-contributor.yml
name: AI Contributor Pipeline

on:
  issues:
    types: [labeled]  # 当 issue 被打上 "ai-take" 标签时触发

jobs:
  generate-pr:
    runs-on: ubuntu-latest
    if: contains(github.event.label.name, 'ai-take')
    steps:
      - name: Checkout repo
        uses: actions/checkout@v4

      - name: Setup Go
        uses: actions/setup-go@v5
        with:
          go-version: '1.22'

      - name: Read issue context
        id: issue
        run: |
          ISSUE_TITLE="${{ github.event.issue.title }}"
          ISSUE_BODY="${{ github.event.issue.body }}"
          echo "title=${ISSUE_TITLE}" >> $GITHUB_OUTPUT
          # 把 issue 内容写入文件供后续步骤读取
          echo "${ISSUE_BODY}" > /tmp/issue_context.txt

      - name: Agent generates patch
        id: agent
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        run: |
          python3 scripts/agent_generate.py \
            --repo "$(pwd)" \
            --issue-file /tmp/issue_context.txt \
            --issue-title "${{ steps.issue.outputs.title }}" \
            --patch-dir /tmp/patch

      - name: Run local tests before PR
        run: |
          cd /tmp/patch/workspace
          go test ./... 2>&1 | tee /tmp/test_results.txt
          # 如果测试失败，不继续提 PR
          if grep -q "FAIL" /tmp/test_results.txt; then
            echo "Tests failed, aborting PR submission"
            exit 1
          fi

      - name: Submit PR
        env:
          GH_TOKEN: ${{ secrets.AI_CONTRIBUTOR_TOKEN }}
        run: |
          cd /tmp/patch/workspace
          BRANCH_NAME="ai-fix-$(date +%Y%m%d-%H%M%S)"
          git checkout -b "$BRANCH_NAME"
          git add -A
          git commit -m "fix: ${{
            steps.issue.outputs.title }}

          Auto-generated by AI contributor agent.
          See linked issue for context."
          git push origin "$BRANCH_NAME"

          gh pr create \
            --title "fix: ${{
              steps.issue.outputs.title }}" \
            --body "## Context
          Linked issue: #${{
            github.event.issue.number }}

          ## Changes
          Generated by AI agent based on issue description.
          Local tests passed. See CI results for full validation.

          ## Verification
          - [x] Local \`go test ./...\` passed
          - [ ] CI pipeline (awaiting results)" \
            --base main \
            --head "$BRANCH_NAME"

#!/usr/bin/env python3
"""Agent 生成补丁的核心逻辑——简化版，可自行扩展"""

import os, sys, json, subprocess
from openai import OpenAI

def read_repo_structure(repo_path: str, max_files: int = 50) -> str:
    """读取仓库关键文件，给 Agent 提供上下文"""
    structure = []
    for root, dirs, files in os.walk(repo_path):
        # 跳过 vendor、.git 等无关目录
        dirs[:] = [d for d in dirs if d not in {'.git', 'vendor', 'node_modules'}]
        for f in files:
            if f.endswith(('.go', '.yaml', '.yml', '.mod', '.sum')):
                full = os.path.join(root, f)
                rel = os.path.relpath(full, repo_path)
                with open(full, 'r', errors='ignore') as fh:
                    content = fh.read(2000)  # 每个文件最多读 2000 字符
                structure.append(f"--- {rel} ---\n{content}")
    # 限制总量，避免超出 token 预算
    return "\n".join(structure[:max_files])

def generate_patch(client: OpenAI, repo_ctx: str, issue: str) -> dict:
    """调用 LLM 生成改动方案"""
    prompt = f"""You are a contributor to a Kubernetes multi-cluster management project.
The project uses Go for the backend.

## Repository context (key files):
{repo_ctx}

## Issue to fix:
{issue}

## Rules:
- Make the smallest possible change that fixes the issue.
- Do not refactor unrelated code.
- Output ONLY the files to change, with their full new content.
- Format as JSON: {{ "files": [{{ "path": "...", "content": "..." }}] }}

Output the patch now."""

    resp = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.2,  # 低温度，减少随机性
    )
    return json.loads(resp.choices[0].message.content)

def apply_patch(patch: dict, repo_path: str, output_dir: str):
    """把 Agent 输出写入工作目录"""
    workspace = os.path.join(output_dir, "workspace")
    # 先复制整个仓库到工作目录
    subprocess.run(["cp", "-r", repo_path, workspace], check=True)
    # 覆盖 Agent 指定的文件
    for file_entry in patch["files"]:
        target = os.path.join(workspace, file_entry["path"])
        os.makedirs(os.path.dirname(target), exist_ok=True)
        with open(target, 'w') as f:
            f.write(file_entry["content"])
    print(f"Patch applied to {workspace}")

def main():
    repo_path = sys.argv[sys.argv.index("--repo") + 1]
    issue_file = sys.argv[sys.argv.index("--issue-file") + 1]
    issue_title = sys.argv[sys.argv.index("--issue-title") + 1]
    patch_dir = sys.argv[sys.argv.index("--patch-dir") + 1]

    with open(issue_file) as f:
        issue_body = f.read()
    issue = f"Title: {issue_title}\nBody: {issue_body}"

    repo_ctx = read_repo_structure(repo_path)
    client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

    patch = generate_patch(client, repo_ctx, issue)
    apply_patch(patch, repo_path, patch_dir)

if __name__ == "__main__":
    main()

# 1. 设置 secrets（在 GitHub 仓库 Settings > Secrets 里添加）
# OPENAI_API_KEY: 你的 OpenAI key
# AI_CONTRIBUTOR_TOKEN: 一个有 repo 写权限的 GitHub token

# 2. 安装依赖
pip install openai

# 3. 在 issue 上打标签触发
# 在 GitHub issue 页面添加标签 "ai-take"，Pipeline 自动启动

# 4. 本地测试脚本（可选，先手动跑一遍验证）
python3 scripts/agent_generate.py \
  --repo ./ \
  --issue-file ./test-issue.txt \
  --issue-title "Fix nil pointer in cluster status handler" \
  --patch-dir /tmp/patch-test
cd /tmp/patch-test/workspace && go test ./...