WuYanBo
/
CodeReview
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add PR synchronization prompt and incorporate commit diffs and conversation history into PR review and chat.

This commit is contained in:
vinland100 2026-01-04 14:34:29 +08:00
parent 23ac263d76
commit d01d3850df
2 changed files with 153 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

128
backend/app/core/ci_prompts.py
View File

@ -10,121 +10,143 @@ from typing import Optional
# Base Template # Base Template
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# strict structure to ensure the LLM has all necessary context without hallucinations. # strict structure to ensure the LLM has all necessary context without hallucinations.
# -----------------------------------------------------------------------------
# 基础模板
# -----------------------------------------------------------------------------
# 严格的结构，确保 LLM 拥有所有必要的上下文而不会产生幻觉。
PROMPT_TEMPLATE = """ PROMPT_TEMPLATE = """
### ROLE ### 角色
{system_prompt} {system_prompt}
### CONTEXT FROM REPOSITORY ### 仓库上下文
The following code snippets were retrieved from the existing repository to provide context: 系统从现有仓库中检索了以下代码片段，以提供背景信息：
{repo_context} {repo_context}
### PR DIFF / CHANGES ### PR 差异 / 变更内容 (Diff)
The following are the actual changes in this Pull Request (or specific commit): 以下是本次 Pull Request (或特定提交) 中的实际变更：
{diff_content} {diff_content}
### CONVERSATION HISTORY ### 对话历史
{conversation_history} {conversation_history}
### TASK ### 任务
{task_description} {task_description}
### OUTPUT FORMAT ### 输出格式
{output_format} {output_format}
""" """
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# 1. PR Review Prompts # 1. PR 评审提示词
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
REVIEW_SYSTEM_PROMPT = """ REVIEW_SYSTEM_PROMPT = """
You are DeepAudit Bot, an expert Senior Security Engineer and Code Reviewer. 你是 AI Code Review Bot，一位资深的安全工程师和代码审查专家。
Your goal is to identify security vulnerabilities, potential bugs, and code quality issues in the provided Pull Request changes. 你的目标是识别提供的 Pull Request 变更中的安全漏洞、潜在 Bug 和代码质量问题。
You must ground your analysis in the provided Repository Context to understand how the changes impact the broader system. 你必须基于提供的“仓库上下文”进行分析，以理解变更对整个系统的影响。
""" """
PR_REVIEW_TASK = """ PR_REVIEW_TASK = """
Analyze the "PR DIFF / CHANGES" above, considering the "CONTEXT FROM REPOSITORY". 分析上方的“PR 差异 / 变更内容 (Diff)”，并结合“仓库上下文”。
1. **Security Analysis**: Identify any security risks (e.g., Injection, Auth bypass, Hardcoded secrets, etc.). 1. **安全分析**：识别任何安全风险（例如：注入、权限绕过、硬编码密钥等）。
2. **Logic & Bugs**: Find edge cases or logic errors introduced in this change. 2. **逻辑与 Bug**：寻找本次变更引入的边界情况或逻辑错误。
3. **Quality & Performance**: Point out maintainability issues or performance bottlenecks. 3. **质量与性能**：指出可维护性问题或性能瓶颈。
4. **Context check**: Use the repo context to verify if function calls or contract changes are valid. 4. **上下文检查**：利用仓库上下文核实函数调用或代码变更是否有效且符合现有架构。
Ignore minor formatting/linting issues unless they severely impact readability. 除非严重影响可读性，否则请忽略细微的格式或 Lint 问题。
""" """
PR_REVIEW_OUTPUT_FORMAT = """ PR_REVIEW_OUTPUT_FORMAT = """
Output ONLY a Markdown response in the following format: 仅输出 Markdown 格式的响应，格式如下：
## 🔍 DeepAudit Review Summary ## 🔍 DeepAudit 评审摘要
<Short summary of the changes and overall risk level> <简要总结变更内容及整体风险等级>
## 🛡️ Key Issues Found ## 🛡️ 发现评审意见
### [Severity: High/Medium/Low] <Title of Issue> ### [严重程度: 高/中/低] <意见标题>
- **File**: `<filepath>` - **文件**: `<文件路径>`
- **Problem**: <Description> - **意见**: <描述具体问题或改进建议>
- **Context**: <Why this is an issue based on repo context> - **上下文**: <基于仓库上下文说明为什么这值得关注>
- **Suggestion**: - **改进建议**:
```<language> ```<语言>
<code fix> <修复建议代码>
``` ```
... (Repeat for other issues) ... (按需重复上述结构)
## 💡 Improvements ## 💡 优化建议
- <Bullet points for minor improvements> - <细微改进的列表>
""" """
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# 2. Incremental (Sync) Review Prompts # 2. 增量 (Sync) 评审提示词
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
PR_SYNC_TASK = """ PR_SYNC_TASK = """
The user has pushed new commits to the existing Pull Request. 用户向现有的 Pull Request 推送了新的提交。
Focus ONLY on the changes in "PR DIFF / CHANGES" (which are the new commits). 请参考下方的“PR 差异 / 变更内容 (Diff)”中的 **全量差异 (Total Diff)** 以了解整个 PR 的背景，
Check if these new changes introduce any new issues or fail to address previous concerns (visible in history). 但请**重点分析并评审**其中的 **本次提交差异 (Recent Sync Diff)**。
1. **安全分析**：识别本次新提交是否引入了任何安全风险。
2. **逻辑与 Bug**：寻找本次新提交中的边界情况或逻辑错误。
3. **回归检查**：核实本次新提交是否解决了之前提到的疑虑，或者是否破坏了已有逻辑。
4. **上下文检查**：利用仓库上下文核实新代码是否有效。
请确保评审意见清晰指出哪些是针对本次新提交的反馈。
""" """
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# 3. Chat / Q&A Prompts # 3. 聊天 / 问答提示词
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
CHAT_SYSTEM_PROMPT = """ CHAT_SYSTEM_PROMPT = """
You are DeepAudit Bot, a helpful AI assistant integrated into the CI/CD workflow. 你是 DeepAudit Bot，一个集成在 CI/CD 工作流中的得力 AI 助手。
You are chatting with a developer in a Pull Request comment thread. 你正在 PR 评论区与开发者交流。
The user has mentioned you (@ai-bot) to ask a question or request clarification. 用户提到了你 (@ai-bot) 以询问问题或请求澄清。
You have access to the relevant snippets of the codebase via RAG (Retrieval Augmented Generation). 你可以通过 RAG (检索增强生成) 访问代码库的相关片段，并能看到当前的 PR 差异。
""" """
BOT_CHAT_TASK = """ BOT_CHAT_TASK = """
Answer the user's question or respond to their comment found in "CONVERSATION HISTORY". 回答用户在“对话历史”中提出的问题或评论。
Use the "CONTEXT FROM REPOSITORY" to provide accurate, specific answers about the code. 利用“仓库上下文”和“PR 差异”来提供关于代码的准确、具体的回答。
If the context doesn't contain the answer, admit it or provide a best-effort answer based on general knowledge. 如果上下文中不包含答案，请如实告知，或基于通用知识提供最佳建议。
Do NOT repeat the user's question. Go straight to the answer. 请不要重复用户的问题，直接开始回答。
""" """
BOT_CHAT_OUTPUT_FORMAT = """ BOT_CHAT_OUTPUT_FORMAT = """
Markdown text. Be concise but technical. Markdown 格式。简洁且具有技术专业性。
""" """
def build_pr_review_prompt(diff: str, context: str, history: str = "None") -> str: def build_pr_review_prompt(diff: str, context: str, history: str = "无") -> str:
return PROMPT_TEMPLATE.format( return PROMPT_TEMPLATE.format(
system_prompt=REVIEW_SYSTEM_PROMPT, system_prompt=REVIEW_SYSTEM_PROMPT,
repo_context=context if context else "No additional context retrieved.", repo_context=context if context else "未检索到相关的仓库上下文。",
diff_content=diff, diff_content=diff,
conversation_history=history, conversation_history=history,
task_description=PR_REVIEW_TASK, task_description=PR_REVIEW_TASK,
output_format=PR_REVIEW_OUTPUT_FORMAT output_format=PR_REVIEW_OUTPUT_FORMAT
) )
def build_chat_prompt(user_query: str, context: str, history: str) -> str: def build_pr_sync_prompt(total_diff: str, sync_diff: str, context: str, history: str) -> str:
# Note: user_query is conceptually part of the history/task combined_diff = f"--- [PR 全量差异 (Total Diff)] ---\n{total_diff}\n\n--- [本次提交差异 (Recent Sync Diff)] ---\n{sync_diff}"
return PROMPT_TEMPLATE.format(
system_prompt=REVIEW_SYSTEM_PROMPT,
repo_context=context if context else "未检索到相关的仓库上下文。",
diff_content=combined_diff,
conversation_history=history,
task_description=PR_SYNC_TASK,
output_format=PR_REVIEW_OUTPUT_FORMAT
)
def build_chat_prompt(user_query: str, context: str, history: str, diff: str = "暂无相关 Diff") -> str:
# 注意：user_query 在概念上是对话历史/任务的一部分
return PROMPT_TEMPLATE.format( return PROMPT_TEMPLATE.format(
system_prompt=CHAT_SYSTEM_PROMPT, system_prompt=CHAT_SYSTEM_PROMPT,
repo_context=context if context else "No additional context retrieved.", repo_context=context if context else "未检索到相关的仓库上下文。",
diff_content="[Not applicable for general chat, unless user refers to recent changes]", diff_content=diff,
conversation_history=history, conversation_history=history,
task_description=BOT_CHAT_TASK + f"\n\nUSER QUESTION: {user_query}", task_description=BOT_CHAT_TASK + f"\n\n用户问题: {user_query}",
output_format=BOT_CHAT_OUTPUT_FORMAT output_format=BOT_CHAT_OUTPUT_FORMAT
) )

91
backend/app/services/ci_service.py
View File

@ -23,6 +23,7 @@ from app.models.project import Project
from app.models.ci import PRReview from app.models.ci import PRReview
from app.core.ci_prompts import ( from app.core.ci_prompts import (
build_pr_review_prompt, build_pr_review_prompt,
build_pr_sync_prompt,
build_chat_prompt, build_chat_prompt,
PR_SYNC_TASK PR_SYNC_TASK
) )
@ -93,13 +94,26 @@ class CIService:
context_results = await retriever.retrieve(diff_text[:1000], top_k=5) context_results = await retriever.retrieve(diff_text[:1000], top_k=5)
repo_context = "\n".join([r.to_context_string() for r in context_results]) repo_context = "\n".join([r.to_context_string() for r in context_results])
# 5. Generate Review # 5. 生成评审
history = ""
if action == "synchronize": if action == "synchronize":
prompt = build_pr_review_prompt(diff_text, repo_context, history) # 增量同步模式：获取全部对话历史
prompt += f"\n\nNOTE: {PR_SYNC_TASK}" history = await self._get_conversation_history(repo, pr_number)
# 获取本次同步的具体差异 (commit diff)
before_sha = payload.get("before")
after_sha = payload.get("after")
sync_diff = ""
if before_sha and after_sha:
sync_diff = await self._get_commit_diff(repo, before_sha, after_sha)
if not sync_diff:
sync_diff = "（无法获取本次提交的具体差异，请参考全量差异）"
prompt = build_pr_sync_prompt(diff_text, sync_diff, repo_context, history)
else: else:
# 新建 PR 模式：历史为空
history = ""
prompt = build_pr_review_prompt(diff_text, repo_context, history) prompt = build_pr_review_prompt(diff_text, repo_context, history)
# Call LLM # Call LLM
@ -229,12 +243,15 @@ class CIService:
repo_context = "\n".join([r.to_context_string() for r in context_results]) repo_context = "\n".join([r.to_context_string() for r in context_results])
# 4. Build Prompt # 4. 获取 PR 差异作为上下文
# Fetch conversation history (simplified: just current comment) diff_text = await self._get_pr_diff(repo, issue.get("number"))
history = f"User: {query}"
prompt = build_chat_prompt(query, repo_context, history) # 5. 构建提示词
# 获取全部 PR 对话历史作为上下文
history = await self._get_conversation_history(repo, issue.get("number"))
prompt = build_chat_prompt(query, repo_context, history, diff=diff_text)
# 5. Generate Answer # 6. 生成回答
response = await self.llm_service.chat_completion_raw( response = await self.llm_service.chat_completion_raw(
messages=[{"role": "user", "content": prompt}], messages=[{"role": "user", "content": prompt}],
temperature=0.4 temperature=0.4
@ -242,10 +259,10 @@ class CIService:
answer = response["content"] answer = response["content"]
# 6. Reply # 7. 回复
# Append context info footer # 附加上下文信息页脚
footer_parts = [f"`{r.file_path}`" for r in context_results] footer_parts = [f"`{r.file_path}`" for r in context_results]
footer = "\n\n---\n*Context used: " + (", ".join(footer_parts) if footer_parts else "None (General knowledge used)") + "*" footer = "\n\n---\n*本次回答参考了以下文件上下文: " + (", ".join(footer_parts) if footer_parts else "无（使用了模型通用知识）") + "*"
await self._post_gitea_comment(repo, issue.get("number"), answer + footer) await self._post_gitea_comment(repo, issue.get("number"), answer + footer)
# 6. Record (Optional, maybe just log) # 6. Record (Optional, maybe just log)
@ -444,6 +461,25 @@ class CIService:
logger.error(f"Failed to fetch PR diff: {e}") logger.error(f"Failed to fetch PR diff: {e}")
return "" return ""
async def _get_commit_diff(self, repo: Dict, before: str, after: str) -> str:
"""
Fetch the diff between two commits from Gitea API
"""
api_url = f"{settings.GITEA_HOST_URL}/api/v1/repos/{repo['owner']['login']}/{repo['name']}/compare/{before}...{after}.diff"
headers = {"Authorization": f"token {settings.GITEA_BOT_TOKEN}"}
try:
async with httpx.AsyncClient() as client:
resp = await client.get(api_url, headers=headers)
if resp.status_code == 200:
return resp.text
else:
logger.error(f"Failed to fetch commit diff: {resp.status_code} - {resp.text[:200]}")
return ""
except Exception as e:
logger.error(f"Failed to fetch commit diff: {e}")
return ""
async def _post_gitea_comment(self, repo: Dict, issue_number: int, body: str): async def _post_gitea_comment(self, repo: Dict, issue_number: int, body: str):
if not settings.GITEA_HOST_URL or not settings.GITEA_BOT_TOKEN: if not settings.GITEA_HOST_URL or not settings.GITEA_BOT_TOKEN:
logger.error("GITEA_HOST_URL or GITEA_BOT_TOKEN not configured") logger.error("GITEA_HOST_URL or GITEA_BOT_TOKEN not configured")
@ -463,3 +499,32 @@ class CIService:
logger.error(f"Gitea API Error: {resp.status_code} - {resp.text}") logger.error(f"Gitea API Error: {resp.status_code} - {resp.text}")
except Exception as e: except Exception as e:
logger.error(f"Failed to post Gitea comment: {e}") logger.error(f"Failed to post Gitea comment: {e}")
async def _get_conversation_history(self, repo: Dict, issue_number: int) -> str:
"""
Fetch the conversation history (comments) from Gitea API
"""
if not settings.GITEA_HOST_URL or not settings.GITEA_BOT_TOKEN:
return "无"
api_url = f"{settings.GITEA_HOST_URL}/api/v1/repos/{repo['owner']['login']}/{repo['name']}/issues/{issue_number}/comments"
headers = {"Authorization": f"token {settings.GITEA_BOT_TOKEN}"}
try:
async with httpx.AsyncClient() as client:
resp = await client.get(api_url, headers=headers)
if resp.status_code == 200:
comments = resp.json()
history_parts = []
for c in comments:
user = c.get("user", {}).get("username") or c.get("user", {}).get("login") or "未知用户"
body = c.get("body", "")
history_parts.append(f"{user}: {body}")
return "\n".join(history_parts) if history_parts else "无"
else:
logger.error(f"Failed to fetch conversation history: {resp.status_code} - {resp.text[:200]}")
return "无"
except Exception as e:
logger.error(f"Failed to fetch PR conversation history: {e}")
return "无"