Skip to content

Commit 0f8c935

Browse files
yestionyestion
committed
Merge branch 'merge-upstream-batch2' into develop
合并上游第二批特性： - PR sunmh207#130: 项目级别的 Prompt 模板配置 - PR sunmh207#148: 批量审查功能 主要功能： 1. 支持为不同项目设置专属的 prompt 模板文件 2. 批量审查：按文件分批次审查，避免上下文过长 3. 智能汇总：多批次审查结果自动汇总 4. 保留 Gitea 平台支持
2 parents 1bfe3cc + cdeff89 commit 0f8c935

3 files changed

Lines changed: 219 additions & 22 deletions

File tree

‎biz/queue/worker.py‎

Lines changed: 20 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -39,8 +39,10 @@ def handle_push_event(webhook_data: dict, gitlab_token: str, gitlab_url: str, gi
3939
review_result = "关注的文件没有修改"
4040

4141
if len(changes) > 0:
42+
project_name = webhook_data['project']['name']
4243
commits_text = ';'.join(commit.get('message', '').strip() for commit in commits)
43-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
44+
code_reviewer = CodeReviewer()
45+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
4446
score = CodeReviewer.parse_review_score(review_text=review_result)
4547
for item in changes:
4648
additions += item['additions']
@@ -133,9 +135,11 @@ def handle_merge_request_event(webhook_data: dict, gitlab_token: str, gitlab_url
133135
logger.error('Failed to get commits')
134136
return
135137

136-
# review 代码
138+
# review 代码 - 使用批量审查方法
139+
project_name = webhook_data['project']['name']
137140
commits_text = ';'.join(commit['title'] for commit in commits)
138-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
141+
code_reviewer = CodeReviewer()
142+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
139143

140144
# 将review结果提交到Gitlab的 notes
141145
handler.add_merge_request_notes(f'Auto Review Result: \n{review_result}')
@@ -189,8 +193,10 @@ def handle_github_push_event(webhook_data: dict, github_token: str, github_url:
189193
review_result = "关注的文件没有修改"
190194

191195
if len(changes) > 0:
196+
project_name = webhook_data['repository']['name']
192197
commits_text = ';'.join(commit.get('message', '').strip() for commit in commits)
193-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
198+
code_reviewer = CodeReviewer()
199+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
194200
score = CodeReviewer.parse_review_score(review_text=review_result)
195201
for item in changes:
196202
additions += item.get('additions', 0)
@@ -273,9 +279,11 @@ def handle_github_pull_request_event(webhook_data: dict, github_token: str, gith
273279
logger.error('Failed to get commits')
274280
return
275281

276-
# review 代码
282+
# review 代码 - 使用批量审查方法
283+
project_name = webhook_data['repository']['name']
277284
commits_text = ';'.join(commit['title'] for commit in commits)
278-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
285+
code_reviewer = CodeReviewer()
286+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
279287

280288
# 将review结果提交到GitHub的 notes
281289
handler.add_pull_request_notes(f'Auto Review Result: \n{review_result}')
@@ -328,8 +336,10 @@ def handle_gitea_push_event(webhook_data: dict, gitea_token: str, gitea_url: str
328336
review_result = "关注的文件没有修改"
329337

330338
if len(changes) > 0:
339+
project_name = webhook_data.get('repository', {}).get('name')
331340
commits_text = ';'.join(commit.get('message', '').strip() for commit in commits)
332-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
341+
code_reviewer = CodeReviewer()
342+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
333343
score = CodeReviewer.parse_review_score(review_text=review_result)
334344
for item in changes:
335345
additions += item.get('additions', 0)
@@ -406,8 +416,10 @@ def handle_gitea_pull_request_event(webhook_data: dict, gitea_token: str, gitea_
406416
logger.error('Failed to get commits for Gitea pull request')
407417
return
408418

419+
project_name = webhook_data.get('repository', {}).get('name')
409420
commits_text = ';'.join(commit.get('title', '') for commit in commits)
410-
review_result = CodeReviewer().review_and_strip_code(str(changes), commits_text)
421+
code_reviewer = CodeReviewer()
422+
review_result = code_reviewer.review_changes_in_batches(changes, commits_text, project_name)
411423

412424
handler.add_pull_request_notes(f'Auto Review Result: \n{review_result}')
413425

‎biz/utils/code_reviewer.py‎

Lines changed: 147 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
import abc
22
import os
33
import re
4-
from typing import Dict, Any, List
4+
from typing import Dict, Any, List, Optional
55

66
import yaml
77
from jinja2 import Template
@@ -16,11 +16,20 @@ class BaseReviewer(abc.ABC):
1616

1717
def __init__(self, prompt_key: str):
1818
self.client = Factory().getClient()
19-
self.prompts = self._load_prompts(prompt_key, os.getenv("REVIEW_STYLE", "professional"))
19+
self.prompts = self._load_prompts(prompt_key)
2020

21-
def _load_prompts(self, prompt_key: str, style="professional") -> Dict[str, Any]:
21+
def _load_prompts(
22+
self, prompt_key: str, style: Optional[str] = None, prompt_templates_file: Optional[str] = None
23+
) -> Dict[str, Any]:
2224
"""加载提示词配置"""
23-
prompt_templates_file = "conf/prompt_templates.yml"
25+
if not style:
26+
# 如果未提供, 从环境变量中获取审查风格，默认为 "professional"
27+
style = os.getenv("REVIEW_STYLE", "professional")
28+
29+
if not prompt_templates_file:
30+
# 如果未提供, 使用默认的提示词配置文件路径
31+
prompt_templates_file = "conf/prompt_templates.yml"
32+
2433
try:
2534
# 在打开 YAML 文件时显式指定编码为 UTF-8，避免使用系统默认的 GBK 编码。
2635
with open(prompt_templates_file, "r", encoding="utf-8") as file:
@@ -60,12 +69,13 @@ class CodeReviewer(BaseReviewer):
6069
def __init__(self):
6170
super().__init__("code_review_prompt")
6271

63-
def review_and_strip_code(self, changes_text: str, commits_text: str = "") -> str:
72+
def review_and_strip_code(self, changes_text: str, commits_text: str = "", project_name: str = "") -> str:
6473
"""
6574
Review判断changes_text超出取前REVIEW_MAX_TOKENS个token，超出则截断changes_text，
6675
调用review_code方法，返回review_result，如果review_result是markdown格式，则去掉头尾的```
67-
:param changes_text:
68-
:param commits_text:
76+
:param changes_text: 代码变更内容
77+
:param commits_text: 提交信息
78+
:param project_name: 项目名称
6979
:return:
7080
"""
7181
# 如果超长，取前REVIEW_MAX_TOKENS个token
@@ -80,29 +90,152 @@ def review_and_strip_code(self, changes_text: str, commits_text: str = "") -> st
8090
if tokens_count > review_max_tokens:
8191
changes_text = truncate_text_by_tokens(changes_text, review_max_tokens)
8292

83-
review_result = self.review_code(changes_text, commits_text).strip()
93+
review_result = self.review_code(changes_text, commits_text, project_name).strip()
8494
if review_result.startswith("```markdown") and review_result.endswith("```"):
8595
return review_result[11:-3].strip()
8696
return review_result
8797

88-
def review_code(self, diffs_text: str, commits_text: str = "") -> str:
98+
def review_code(self, diffs_text: str, commits_text: str = "", project_name: str = "") -> str:
8999
"""Review 代码并返回结果"""
100+
normalized_project_name = project_name.replace("-", "_") if project_name else project_name
101+
project_prompts_path = os.getenv(f"{normalized_project_name.upper()}_PROMPT", None)
102+
103+
# 按需重新加载 prompts 配置， 同时也可以支持项目级别提示词的热加载
104+
prompts = (
105+
self._load_prompts(prompt_key="code_review_prompt", prompt_templates_file=project_prompts_path)
106+
if project_prompts_path
107+
else self.prompts
108+
)
90109
messages = [
91-
self.prompts["system_message"],
110+
prompts["system_message"],
92111
{
93112
"role": "user",
94-
"content": self.prompts["user_message"]["content"].format(
95-
diffs_text=diffs_text, commits_text=commits_text
96-
),
113+
"content": prompts["user_message"]["content"].format(diffs_text=diffs_text, commits_text=commits_text),
97114
},
98115
]
99116
return self.call_llm(messages)
100117

118+
def review_changes_in_batches(self, changes: List[Dict[str, Any]], commits_text: str = "", project_name: str = "") -> str:
119+
"""
120+
按文件批次审查代码变更，然后汇总所有审查结果
121+
:param changes: 代码变更列表，每个元素是一个包含文件信息的字典
122+
:param commits_text: 提交信息
123+
:param project_name: 项目名称
124+
:return: 汇总后的审查结果
125+
"""
126+
if not changes:
127+
logger.info("代码变更为空")
128+
return "代码为空"
129+
130+
# 检查是否启用批量审查
131+
batch_review_enabled = os.getenv("BATCH_REVIEW_ENABLED", "1") == "1"
132+
133+
# 如果未启用批量审查，使用原有的一次性审查方式
134+
if not batch_review_enabled:
135+
logger.info("批量审查功能未启用，使用传统一次性审查方式")
136+
return self.review_and_strip_code(str(changes), commits_text, project_name)
137+
138+
review_max_tokens = int(os.getenv("REVIEW_MAX_TOKENS", 10000))
139+
# 获取每批次审查的文件数量配置
140+
files_per_batch = int(os.getenv("BATCH_REVIEW_FILES_PER_BATCH", 1))
141+
logger.info(f"批量审查已启用，每批次审查 {files_per_batch} 个文件")
142+
143+
partial_reviews = []
144+
total_files = len(changes)
145+
146+
# 按配置的批次大小分批进行审查
147+
for batch_start in range(0, total_files, files_per_batch):
148+
batch_end = min(batch_start + files_per_batch, total_files)
149+
batch_changes = changes[batch_start:batch_end]
150+
batch_num = (batch_start // files_per_batch) + 1
151+
total_batches = (total_files + files_per_batch - 1) // files_per_batch
152+
153+
logger.info(f"正在审查第 {batch_num}/{total_batches} 批次 (文件 {batch_start + 1}-{batch_end}/{total_files})")
154+
155+
# 收集当前批次的文件路径
156+
batch_file_paths = [
157+
change.get('new_path') or change.get('old_path', 'unknown')
158+
for change in batch_changes
159+
]
160+
161+
# 将批次内的文件转换为文本
162+
batch_text = str(batch_changes)
163+
164+
# 计算tokens数量，如果超过限制则截断
165+
tokens_count = count_tokens(batch_text)
166+
if tokens_count > review_max_tokens:
167+
logger.warning(f"批次 {batch_num} 的变更超过 {review_max_tokens} tokens，将截断")
168+
batch_text = truncate_text_by_tokens(batch_text, review_max_tokens)
169+
170+
# 审查当前批次，传递 project_name 参数
171+
try:
172+
review_result = self.review_code(batch_text, commits_text, project_name).strip()
173+
if review_result.startswith("```markdown") and review_result.endswith("```"):
174+
review_result = review_result[11:-3].strip()
175+
176+
# 添加批次标识
177+
batch_header = f"### 批次 {batch_num} (文件: {', '.join(batch_file_paths)})\n"
178+
partial_reviews.append(f"{batch_header}{review_result}")
179+
logger.info(f"批次 {batch_num} 审查完成")
180+
except Exception as e:
181+
logger.error(f"审查批次 {batch_num} 时出错: {e}")
182+
partial_reviews.append(f"### 批次 {batch_num}\n审查失败: {str(e)}")
183+
184+
# 如果只有一个批次，直接返回结果（去掉批次标识）
185+
if len(partial_reviews) == 1:
186+
# 去掉批次标题行
187+
result = partial_reviews[0]
188+
lines = result.split('\n', 1)
189+
return lines[1] if len(lines) > 1 else result
190+
191+
# 汇总多个批次的审查结果
192+
logger.info(f"开始汇总 {len(partial_reviews)} 个批次的审查结果")
193+
summary_result = self._summarize_reviews(partial_reviews, project_name)
194+
return summary_result
195+
196+
def _summarize_reviews(self, partial_reviews: List[str], project_name: str = "") -> str:
197+
"""
198+
使用 summary_merge_review_prompt 汇总多个审查结果
199+
:param partial_reviews: 各批次的审查结果列表
200+
:param project_name: 项目名称
201+
:return: 汇总后的总审查报告
202+
"""
203+
# 加载汇总提示词，支持项目级别的自定义
204+
normalized_project_name = project_name.replace("-", "_") if project_name else project_name
205+
project_prompts_path = os.getenv(f"{normalized_project_name.upper()}_PROMPT", None)
206+
207+
summary_prompts = (
208+
self._load_prompts(prompt_key="summary_merge_review_prompt", prompt_templates_file=project_prompts_path)
209+
if project_prompts_path
210+
else self._load_prompts("summary_merge_review_prompt", os.getenv("REVIEW_STYLE", "professional"))
211+
)
212+
213+
# 拼接所有分批审查结果
214+
partial_reviews_text = "\n\n---\n\n".join(partial_reviews)
215+
216+
# 构建汇总请求消息
217+
messages = [
218+
summary_prompts["system_message"],
219+
{
220+
"role": "user",
221+
"content": summary_prompts["user_message"]["content"].format(
222+
partial_reviews_text=partial_reviews_text
223+
),
224+
},
225+
]
226+
227+
# 调用LLM进行汇总
228+
summary_result = self.call_llm(messages).strip()
229+
if summary_result.startswith("```markdown") and summary_result.endswith("```"):
230+
summary_result = summary_result[11:-3].strip()
231+
232+
logger.info("审查结果汇总完成")
233+
return summary_result
234+
101235
@staticmethod
102236
def parse_review_score(review_text: str) -> int:
103237
"""解析 AI 返回的 Review 结果，返回评分"""
104238
if not review_text:
105239
return 0
106240
match = re.search(r"总分[:：]\s*(\d+)分?", review_text)
107241
return int(match.group(1)) if match else 0
108-

‎conf/prompt_templates.yml‎

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -41,3 +41,55 @@ code_review_prompt:
4141
4242
提交历史(commits)：
4343
{commits_text}
44+
45+
summary_merge_review_prompt:
46+
system_prompt: |-
47+
你是一位高级软件架构师，现在需要对多个分批完成的代码审查结果进行整合成一个完整的总审查报告。
48+
49+
你的职责：
50+
1. 重新整合多个批次的审查结果，形成"统一评分的总报告"
51+
2. 不得丢失开发者定位问题所需的"批次级详细描述"
52+
3. 在顶层总结问题趋势及关键风险
53+
4. 根据全量问题重新统一打分
54+
55+
---
56+
57+
汇总结构必须包含以下 3 部分：
58+
59+
### 第一部分：全局问题总结与优化建议（进行归类整合，去重问题）
60+
- 从所有批次报告中抽取共性问题进行分类总结
61+
- 以整体角度提出优化方向，而不是重复粘贴批次内容
62+
63+
### 第二部分：分批次详细问题保留区（必须原样结构化保留）
64+
你的任务是按以下格式保留批次细节，不得简化或省略：
65+
```
66+
#### 批次 X（文件范围/来源说明）
67+
<保留该批完整的"问题描述与评分明细"，不得删减内容>
68+
```
69+
70+
这样开发者能快速找到"哪个文件在哪个批次出了什么问题"。
71+
72+
### 第三部分：统一评分明细与总分（你必须重新评分）
73+
- 你需要结合多个批次的影响范围重新量化总评分，而不是平均或取最大值
74+
- 格式如下：
75+
```
76+
- 功能实现的正确性与健壮性：XX分
77+
- 安全性与潜在风险：XX分
78+
- 是否符合最佳实践：XX分
79+
- 性能与资源效率：XX分
80+
- 提交信息清晰性与准确性：XX分
81+
```
82+
83+
最后一行必须为：**总分:XX分**
84+
85+
---
86+
87+
自检规则：
88+
- 若未包含"分批次详细问题保留区"则需重新生成
89+
- 若未重新统一评分，而直接引用批次数值，则需重新评分
90+
- 若缺少"总分:XX分"，必须重新生成
91+
- 若全局总结部分只是重复批次内容，必须进行整合后重新生成
92+
93+
user_prompt: |-
94+
以下是分批次代码审查结果，请将其整合为一个完整的总审查报告，并统一量化评分：
95+
{partial_reviews_text}

0 commit comments

Comments
 (0)