1. بررسی کد آخر شب
ساعت 2 بامداد است
شما ساعت ها در حال رفع اشکال بوده اید. عملکرد درست به نظر می رسد، اما چیزی اشتباه است. شما این احساس را می‌دانید - زمانی که کد باید کار کند اما این کار را نمی‌کند، و دیگر نمی‌توانید دلیل آن را ببینید زیرا مدت زیادی به آن خیره شده‌اید.
def dfs_search_v1(graph, start, target):
"""Find if target is reachable from start."""
visited = set()
stack = start # Looks innocent enough...
while stack:
current = stack.pop()
if current == target:
return True
if current not in visited:
visited.add(current)
for neighbor in graph[current]:
if neighbor not in visited:
stack.append(neighbor)
return False
سفر توسعه دهنده هوش مصنوعی
اگر در حال خواندن این مطلب هستید، احتمالاً تحولی را که هوش مصنوعی در کدنویسی به ارمغان می آورد، تجربه کرده اید. ابزارهایی مانند Gemini Code Assist ، Claude Code و Cursor نحوه نوشتن کد را تغییر داده اند. آنها برای تولید دیگ بخار، پیشنهاد پیاده سازی و تسریع توسعه فوق العاده هستند.
اما شما اینجا هستید زیرا می خواهید عمیق تر بروید. شما می خواهید بدانید که چگونه این سیستم های هوش مصنوعی را بسازید ، نه اینکه فقط از آنها استفاده کنید. شما می خواهید چیزی بسازید که:
- رفتار قابل پیش بینی و قابل ردیابی دارد
- می توان با اطمینان به تولید مستقر شد
- نتایج ثابتی را ارائه می دهد که می توانید به آنها اعتماد کنید
- دقیقاً به شما نشان می دهد که چگونه تصمیم می گیرد
از مصرف کننده تا خالق
امروز، از استفاده از ابزارهای هوش مصنوعی به ساخت آنها جهش خواهید کرد. شما یک سیستم چند عاملی خواهید ساخت که:
- ساختار کد را به طور قطعی تجزیه و تحلیل می کند
- آزمایش های واقعی را برای تأیید رفتار انجام می دهد
- انطباق سبک با لینت های واقعی را تأیید می کند
- یافته ها را به بازخورد عملی ترکیب می کند
- با قابلیت مشاهده کامل در Google Cloud مستقر می شود
2. اولین استقرار عامل شما
سوال توسعه دهنده
"من LLM ها را درک می کنم، از API ها استفاده کرده ام، اما چگونه می توانم از یک اسکریپت پایتون به یک عامل هوش مصنوعی تولیدی بروم؟"
بیایید با تنظیم صحیح محیط خود به این پاسخ پاسخ دهیم، سپس یک عامل ساده برای درک اصول اولیه قبل از فرو رفتن در الگوهای تولید بسازیم.
ابتدا راه اندازی ضروری
قبل از ایجاد هر گونه عاملی، مطمئن شویم که محیط Google Cloud شما آماده است.
به اعتبار Google Cloud نیاز دارید؟
روی Activate Cloud Shell در بالای کنسول Google Cloud کلیک کنید (این نماد شکل ترمینال در بالای صفحه Cloud Shell است)،
شناسه پروژه Google Cloud خود را پیدا کنید:
- Google Cloud Console را باز کنید: https://console.cloud.google.com
- پروژه ای را که می خواهید برای این کارگاه استفاده کنید از منوی کشویی پروژه در بالای صفحه انتخاب کنید.
- شناسه پروژه شما در کارت اطلاعات پروژه در داشبورد نمایش داده می شود
مرحله 1: شناسه پروژه خود را تنظیم کنید
در Cloud Shell، ابزار خط فرمان gcloud قبلاً پیکربندی شده است. دستور زیر را برای تنظیم پروژه فعال خود اجرا کنید. این از متغیر محیطی $GOOGLE_CLOUD_PROJECT استفاده می کند که به طور خودکار در جلسه Cloud Shell برای شما تنظیم می شود.
gcloud config set project $GOOGLE_CLOUD_PROJECT
مرحله 2: تنظیمات خود را تأیید کنید
در مرحله بعد، دستورات زیر را اجرا کنید تا تأیید کنید که پروژه شما به درستی تنظیم شده است و احراز هویت شما تأیید شده است.
# Confirm project is set
echo "Current project: $(gcloud config get-value project)"
# Check authentication status
gcloud auth list
باید ID پروژه خود را چاپ شده ببینید و حساب کاربری شما با (ACTIVE) در کنار آن فهرست شده است.
اگر حساب شما به عنوان فعال لیست نشده است، یا اگر با خطای احراز هویت مواجه شدید، دستور زیر را اجرا کنید تا وارد شوید:
gcloud auth application-default login
مرحله 3: API های ضروری را فعال کنید
ما حداقل به این APIها برای عامل اصلی نیاز داریم:
gcloud services enable \
aiplatform.googleapis.com \
compute.googleapis.com
این ممکن است یک یا دو دقیقه طول بکشد. خواهید دید:
Operation "operations/..." finished successfully.
مرحله 4: ADK را نصب کنید
# Install the ADK CLI
pip install google-adk --upgrade
# Verify installation
adk --version
شما باید یک شماره نسخه مانند 1.15.0 یا بالاتر را ببینید.
اکنون عامل اصلی خود را ایجاد کنید
با محیط آماده، بیایید آن عامل ساده را ایجاد کنیم.
مرحله 5: از ADK Create استفاده کنید
adk create my_first_agent
دستورات تعاملی را دنبال کنید:
Choose a model for the root agent:
1. gemini-2.5-flash
2. Other models (fill later)
Choose model (1, 2): 1
1. Google AI
2. Vertex AI
Choose a backend (1, 2): 2
Enter Google Cloud project ID [auto-detected-from-gcloud]:
Enter Google Cloud region [us-central1]:
مرحله 6: آنچه ایجاد شده را بررسی کنید
cd my_first_agent
ls -la
سه فایل پیدا خواهید کرد:
.env # Configuration (auto-populated with your project)
__init__.py # Package marker
agent.py # Your agent definition
مرحله 7: بررسی سریع پیکربندی
# Verify the .env was created correctly
cat .env
# Should show something like:
# GOOGLE_CLOUD_PROJECT=your-project-id
# GOOGLE_CLOUD_LOCATION=us-central1
# GOOGLE_GENAI_USE_VERTEXAI=1
اگر شناسه پروژه گم یا نادرست است، فایل .env را ویرایش کنید:
nano .env # or use your preferred editor
مرحله 8: به کد عامل نگاه کنید
cat agent.py
from google.adk.agents.llm_agent import Agent
root_agent = Agent(
model='gemini-2.5-flash',
name='root_agent',
description='A helpful assistant for user questions.',
instruction='Answer user questions to the best of your knowledge',
)
ساده، تمیز، مینیمال. این "سلام جهان" ماموران شماست.
عامل اصلی خود را آزمایش کنید
مرحله 9: نماینده خود را اجرا کنید
cd ..
adk run my_first_agent
شما باید چیزی شبیه به:
Log setup complete: /tmp/agents_log/agent.20250930_162430.log
To access latest log: tail -F /tmp/agents_log/agent.latest.log
[user]:
مرحله 10: برخی از پرس و جوها را امتحان کنید
در ترمینالی که adk run اجرا می شود، یک اعلان خواهید دید. سوالات خود را تایپ کنید:
[user]: Hey, what can you do?
[root_agent]: I can help answer questions, generate text, explain concepts, assist with problem-solving, and provide information on a wide range of topics. What would you like to know or discuss?
[user]: What's the current weather in Mountain View?
[root_agent]: I'm sorry, but I cannot provide real-time weather information as I don't have access to current data or weather services. To get the current weather in Mountain View, I recommend checking a weather website or app like weather.com or using a search engine.
به محدودیت توجه کنید - نمی تواند به داده های فعلی دسترسی پیدا کند. بیایید بیشتر فشار بیاوریم:
[user]: Review this Python code: def add(a, b): return a + b
[root_agent]: This is a simple Python function that adds two numbers. Here's my review:
**Strengths:**
- Clean and concise implementation
- Follows Python conventions for simple operations
- Clear function name that describes its purpose
**Suggestions for improvement:**
- Add a docstring to document the function
- Consider adding type hints for better code clarity
- You might want to handle edge cases
Here's an enhanced version:
def add(a: float, b: float) -> float:
"""Add two numbers and return their sum."""
return a + b
عامل می تواند درباره کد بحث کند، اما می تواند:
- آیا برای درک ساختار AST را تجزیه می کنید؟
- تست هایی را اجرا کنید تا مطمئن شوید کار می کند؟
- مطابقت سبک را بررسی کنید؟
- بررسی های قبلی خود را به خاطر دارید؟
نه. اینجاست که ما به معماری نیاز داریم.
🏃🚪 خروج با
Ctrl+C
وقتی کاوش تمام شد
3. آماده سازی محل کار تولید خود
راه حل: معماری آماده تولید
این عامل ساده نقطه شروع را نشان داد، اما یک سیستم تولید به ساختار قوی نیاز دارد. ما اکنون یک پروژه کامل را راه اندازی می کنیم که اصول تولید را در بر می گیرد.
راه اندازی بنیاد
شما قبلاً پروژه Google Cloud خود را برای عامل اصلی پیکربندی کرده اید. حال بیایید فضای کاری کامل تولید را با تمام ابزارها، الگوها و زیرساخت های مورد نیاز برای یک سیستم واقعی آماده کنیم.
مرحله 1: پروژه ساخت یافته را دریافت کنید
ابتدا از هر adk run با Ctrl+C خارج شده و پاکسازی کنید:
# Clean up the basic agent
cd ~ # Make sure you're not inside my_first_agent
rm -rf my_first_agent
# Get the production scaffold
git clone https://github.com/ayoisio/adk-code-review-assistant.git
cd adk-code-review-assistant
git checkout codelab
مرحله 2: ایجاد و فعال کردن محیط مجازی
# Create the virtual environment
python -m venv .venv
# Activate it
# On macOS/Linux:
source .venv/bin/activate
# On Windows:
# .venv\Scripts\activate
تأیید : اکنون درخواست شما باید (.venv) را در ابتدا نشان دهد.
مرحله 3: Dependencies را نصب کنید
pip install -r code_review_assistant/requirements.txt
# Install the package in editable mode (enables imports)
pip install -e .
این نصب می کند:
-
google-adk- چارچوب ADK -
pycodestyle- برای بررسی PEP 8 -
vertexai- برای استقرار ابر - سایر وابستگی های تولید
پرچم -e به شما امکان می دهد ماژول های code_review_assistant از هر جایی وارد کنید.
مرحله 4: محیط خود را پیکربندی کنید
# Copy the example environment file
cp .env.example .env
# Edit .env and replace the placeholders:
# - GOOGLE_CLOUD_PROJECT=your-project-id → your actual project ID
# - Keep other defaults as-is
تأیید : پیکربندی خود را بررسی کنید:
cat .env
باید نشان دهد:
GOOGLE_CLOUD_PROJECT=your-actual-project-id
GOOGLE_CLOUD_LOCATION=us-central1
GOOGLE_GENAI_USE_VERTEXAI=TRUE
مرحله 5: از احراز هویت اطمینان حاصل کنید
از آنجایی که قبلاً gcloud auth اجرا کرده اید، اجازه دهید فقط تأیید کنیم:
# Check current authentication
gcloud auth list
# Should show your account with (ACTIVE)
# If not, run:
gcloud auth application-default login
مرحله 6: API های تولید اضافی را فعال کنید
ما قبلاً APIهای اصلی را فعال کرده‌ایم. حالا موارد تولیدی را اضافه کنید:
gcloud services enable \
sqladmin.googleapis.com \
run.googleapis.com \
cloudbuild.googleapis.com \
artifactregistry.googleapis.com \
storage.googleapis.com \
cloudtrace.googleapis.com
این امکان را فراهم می کند:
- SQL Admin : برای Cloud SQL در صورت استفاده از Cloud Run
- Cloud Run : برای استقرار بدون سرور
- Cloud Build : برای استقرار خودکار
- رجیستری مصنوع : برای تصاویر کانتینر
- ذخیره سازی ابری : برای مصنوعات و صحنه سازی
- Cloud Trace : برای قابلیت مشاهده
مرحله 7: مخزن رجیستری مصنوع را ایجاد کنید
استقرار ما تصاویر کانتینری خواهد ساخت که نیاز به خانه دارند:
gcloud artifacts repositories create code-review-assistant-repo \
--repository-format=docker \
--location=us-central1 \
--description="Docker repository for Code Review Assistant"
باید ببینید:
Created repository [code-review-assistant-repo].
اگر از قبل وجود داشته باشد (شاید از تلاش قبلی)، خوب است - یک پیام خطایی خواهید دید که می توانید نادیده بگیرید.
مرحله 8: مجوز IAM را اعطا کنید
# Get your project number
PROJECT_NUMBER=$(gcloud projects describe $GOOGLE_CLOUD_PROJECT \
--format="value(projectNumber)")
# Define the service account
SERVICE_ACCOUNT="${PROJECT_NUMBER}@cloudbuild.gserviceaccount.com"
# Grant necessary roles
gcloud projects add-iam-policy-binding $GOOGLE_CLOUD_PROJECT \
--member="serviceAccount:${SERVICE_ACCOUNT}" \
--role="roles/run.admin"
gcloud projects add-iam-policy-binding $GOOGLE_CLOUD_PROJECT \
--member="serviceAccount:${SERVICE_ACCOUNT}" \
--role="roles/iam.serviceAccountUser"
gcloud projects add-iam-policy-binding $GOOGLE_CLOUD_PROJECT \
--member="serviceAccount:${SERVICE_ACCOUNT}" \
--role="roles/cloudsql.admin"
gcloud projects add-iam-policy-binding $GOOGLE_CLOUD_PROJECT \
--member="serviceAccount:${SERVICE_ACCOUNT}" \
--role="roles/storage.admin"
هر فرمان خروجی خواهد داشت:
Updated IAM policy for project [your-project-id].
آنچه شما انجام داده اید
اکنون فضای کاری تولید شما کاملاً آماده است:
✅ پروژه Google Cloud پیکربندی و احراز هویت شد
✅ عامل پایه برای درک محدودیت ها آزمایش شده است
✅ کد پروژه با متغیرهای استراتژیک آماده است
✅ وابستگی های ایزوله شده در محیط مجازی
✅ همه API های لازم فعال هستند
✅ رجیستری کانتینر آماده استقرار
✅ مجوزهای IAM به درستی پیکربندی شده اند
✅ متغیرهای محیطی به درستی تنظیم شده اند
اکنون شما آماده ساختن یک سیستم هوش مصنوعی واقعی با ابزارهای قطعی، مدیریت حالت و معماری مناسب هستید.
4. ساختن اولین نماینده خود
چه چیزی ابزارها را از LLM متفاوت می کند
وقتی از یک LLM می‌پرسید «چند تابع در این کد وجود دارد؟»، از تطبیق و تخمین الگو استفاده می‌کند. وقتی از ابزاری استفاده می‌کنید که ast.parse() پایتون را فراخوانی می‌کند، درخت نحو واقعی را تجزیه می‌کند - حدس نمی‌زنید، هر بار نتیجه یکسان است.
این بخش ابزاری را ایجاد می کند که ساختار کد را به طور قطعی تجزیه و تحلیل می کند، سپس آن را به عاملی متصل می کند که می داند چه زمانی باید آن را فراخوانی کند.
مرحله 1: شناخت داربست
بیایید ساختاری را که پر می‌کنید بررسی کنیم.
👉 باز کنید
code_review_assistant/tools.py
تابع analyze_code_structure را با نظرات نگهدارنده مکان مشاهده خواهید کرد که کجا کد را اضافه می کنید. این تابع در حال حاضر ساختار اصلی را دارد - شما آن را گام به گام بهبود خواهید داد.
مرحله 2: فضای ذخیره سازی را اضافه کنید
ذخیره سازی حالت به سایر عوامل در خط لوله اجازه می دهد تا بدون اجرای مجدد تجزیه و تحلیل به نتایج ابزار شما دسترسی داشته باشند.
👉 پیدا کنید:
# MODULE_4_STEP_2_ADD_STATE_STORAGE
👉 آن خط واحد را با:
# Store code and analysis for other agents to access
tool_context.state[StateKeys.CODE_TO_REVIEW] = code
tool_context.state[StateKeys.CODE_ANALYSIS] = analysis
tool_context.state[StateKeys.CODE_LINE_COUNT] = len(code.splitlines())
مرحله 3: تجزیه Async را با Thread Pools اضافه کنید
ابزار ما نیاز به تجزیه AST بدون مسدود کردن سایر عملیات دارد. بیایید اجرای async را با Thread Pool اضافه کنیم.
👉 پیدا کنید:
# MODULE_4_STEP_3_ADD_ASYNC
👉 آن خط واحد را با:
# Parse in thread pool to avoid blocking the event loop
loop = asyncio.get_event_loop()
with ThreadPoolExecutor() as executor:
tree = await loop.run_in_executor(executor, ast.parse, code)
مرحله 4: استخراج اطلاعات جامع
حالا بیایید کلاس‌ها، واردات و معیارهای دقیق را استخراج کنیم - همه چیزهایی که برای بررسی کامل کد نیاز داریم.
👉 پیدا کنید:
# MODULE_4_STEP_4_EXTRACT_DETAILS
👉 آن خط واحد را با:
# Extract comprehensive structural information
analysis = await loop.run_in_executor(
executor, _extract_code_structure, tree, code
)
👉 بررسی کنید: عملکرد
analyze_code_structure
در
tools.py
دارای بدنه مرکزی به شکل زیر است:
# Parse in thread pool to avoid blocking the event loop
loop = asyncio.get_event_loop()
with ThreadPoolExecutor() as executor:
tree = await loop.run_in_executor(executor, ast.parse, code)
# Extract comprehensive structural information
analysis = await loop.run_in_executor(
executor, _extract_code_structure, tree, code
)
# Store code and analysis for other agents to access
tool_context.state[StateKeys.CODE_TO_REVIEW] = code
tool_context.state[StateKeys.CODE_ANALYSIS] = analysis
tool_context.state[StateKeys.CODE_LINE_COUNT] = len(code.splitlines())
👉 حالا به پایین صفحه بروید
tools.py
و پیدا کنید:
# MODULE_4_STEP_4_HELPER_FUNCTION
👉 آن خط واحد را با تابع کمکی کامل جایگزین کنید:
def _extract_code_structure(tree: ast.AST, code: str) -> Dict[str, Any]:
"""
Helper function to extract structural information from AST.
Runs in thread pool for CPU-bound work.
"""
functions = []
classes = []
imports = []
docstrings = []
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef):
func_info = {
'name': node.name,
'args': [arg.arg for arg in node.args.args],
'lineno': node.lineno,
'has_docstring': ast.get_docstring(node) is not None,
'is_async': isinstance(node, ast.AsyncFunctionDef),
'decorators': [d.id for d in node.decorator_list
if isinstance(d, ast.Name)]
}
functions.append(func_info)
if func_info['has_docstring']:
docstrings.append(f"{node.name}: {ast.get_docstring(node)[:50]}...")
elif isinstance(node, ast.ClassDef):
methods = []
for item in node.body:
if isinstance(item, ast.FunctionDef):
methods.append(item.name)
class_info = {
'name': node.name,
'lineno': node.lineno,
'methods': methods,
'has_docstring': ast.get_docstring(node) is not None,
'base_classes': [base.id for base in node.bases
if isinstance(base, ast.Name)]
}
classes.append(class_info)
elif isinstance(node, ast.Import):
for alias in node.names:
imports.append({
'module': alias.name,
'alias': alias.asname,
'type': 'import'
})
elif isinstance(node, ast.ImportFrom):
imports.append({
'module': node.module or '',
'names': [alias.name for alias in node.names],
'type': 'from_import',
'level': node.level
})
return {
'functions': functions,
'classes': classes,
'imports': imports,
'docstrings': docstrings,
'metrics': {
'line_count': len(code.splitlines()),
'function_count': len(functions),
'class_count': len(classes),
'import_count': len(imports),
'has_main': any(f['name'] == 'main' for f in functions),
'has_if_main': '__main__' in code,
'avg_function_length': _calculate_avg_function_length(tree)
}
}
def _calculate_avg_function_length(tree: ast.AST) -> float:
"""Calculate average function length in lines."""
function_lengths = []
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef):
if hasattr(node, 'end_lineno') and hasattr(node, 'lineno'):
length = node.end_lineno - node.lineno + 1
function_lengths.append(length)
if function_lengths:
return sum(function_lengths) / len(function_lengths)
return 0.0
مرحله 5: به یک Agent متصل شوید
اکنون ابزار را به عاملی متصل می کنیم که می داند چه زمانی از آن استفاده کند و چگونه نتایج آن را تفسیر کند.
👉 باز کنید
code_review_assistant/sub_agents/review_pipeline/code_analyzer.py
👉 پیدا کنید:
# MODULE_4_STEP_5_CREATE_AGENT
👉 آن خط واحد را با عامل تولید کامل جایگزین کنید:
code_analyzer_agent = Agent(
name="CodeAnalyzer",
model=config.worker_model,
description="Analyzes Python code structure and identifies components",
instruction="""You are a code analysis specialist responsible for understanding code structure.
Your task:
1. Take the code submitted by the user (it will be provided in the user message)
2. Use the analyze_code_structure tool to parse and analyze it
3. Pass the EXACT code to your tool - do not modify, fix, or "improve" it
4. Identify all functions, classes, imports, and structural patterns
5. Note any syntax errors or structural issues
6. Store the analysis in state for other agents to use
CRITICAL:
- Pass the code EXACTLY as provided to the analyze_code_structure tool
- Do not fix syntax errors, even if obvious
- Do not add missing imports or fix indentation
- The goal is to analyze what IS there, not what SHOULD be there
When calling the tool, pass the code as a string to the 'code' parameter.
If the analysis fails due to syntax errors, clearly report the error location and type.
Provide a clear summary including:
- Number of functions and classes found
- Key structural observations
- Any syntax errors or issues detected
- Overall code organization assessment""",
tools=[FunctionTool(func=analyze_code_structure)],
output_key="structure_analysis_summary"
)
تحلیلگر کد خود را تست کنید
اکنون بررسی کنید که آنالایزر شما به درستی کار می کند.
👉 اسکریپت تست را اجرا کنید:
python tests/test_code_analyzer.py
اسکریپت آزمایشی به طور خودکار پیکربندی را از فایل .env شما با استفاده از python-dotenv بارگیری می کند، بنابراین نیازی به تنظیم دستی متغیر محیط نیست.
خروجی مورد انتظار:
INFO:code_review_assistant.config:Code Review Assistant Configuration Loaded:
INFO:code_review_assistant.config: - GCP Project: your-project-id
INFO:code_review_assistant.config: - Artifact Bucket: gs://your-project-artifacts
INFO:code_review_assistant.config: - Models: worker=gemini-2.5-flash, critic=gemini-2.5-pro
Testing code analyzer...
INFO:code_review_assistant.tools:Tool: Analysis complete - 2 functions, 1 classes
=== Analyzer Response ===
The analysis of the provided code shows the following:
* **Functions Found:** 2
* `add(a, b)`: A global function at line 2.
* `multiply(self, x, y)`: A method within the `Calculator` class.
* **Classes Found:** 1
* `Calculator`: A class defined at line 5. Contains one method, `multiply`.
* **Imports:** 0
* **Structural Patterns:** The code defines one global function and one class
with a single method. Both are simple, each with a single return statement.
* **Syntax Errors/Issues:** No syntax errors detected.
* **Overall Code Organization:** The code is well-organized for its small size,
clearly defining a function and a class with a method.
تازه چه اتفاقی افتاد:
- اسکریپت آزمایشی پیکربندی
.envشما را به صورت خودکار بارگیری کرد - ابزار
analyze_code_structure()کد را با استفاده از AST پایتون تجزیه کرد - کمک کننده
_extract_code_structure()توابع، کلاس ها و متریک ها را استخراج کرد - نتایج در حالت جلسه با استفاده از ثابت های
StateKeysذخیره شدند - عامل Code Analyzer نتایج را تفسیر کرد و خلاصه ای را ارائه کرد
عیب یابی:
- "هیچ ماژولی با نام "code_review_assistant" وجود ندارد :
pip install -e .از ریشه پروژه - "آگومان ورودی های کلیدی از دست رفته" : بررسی کنید
.envشما دارایGOOGLE_CLOUD_PROJECT،GOOGLE_CLOUD_LOCATION، وGOOGLE_GENAI_USE_VERTEXAI=true
آنچه شما ساخته اید
اکنون یک تحلیلگر کد آماده تولید دارید که:
✅ AST واقعی پایتون را تجزیه می‌کند - قطعی، نه تطبیق الگو
✅ نتایج را در حالت ذخیره می کند - سایر عوامل می توانند به تجزیه و تحلیل دسترسی داشته باشند
✅ به صورت ناهمزمان اجرا می شود - ابزارهای دیگر را مسدود نمی کند
✅ اطلاعات جامع را استخراج می کند - توابع، کلاس ها، واردات، معیارها
✅ خطاها را به خوبی مدیریت می کند - خطاهای نحوی را با شماره خطوط گزارش می دهد
✅ به یک نماینده متصل می شود - LLM می داند چه زمانی و چگونه از آن استفاده کند
تسلط بر مفاهیم کلیدی
ابزارها در مقابل نمایندگان:
- ابزارها کار قطعی انجام می دهند (تجزیه AST)
- عوامل تصمیم می گیرند که چه زمانی از ابزارها استفاده کنند و نتایج را تفسیر کنند
ارزش بازگشتی در برابر وضعیت:
- بازگشت: چیزی که LLM بلافاصله می بیند
- حالت: آنچه برای سایر عوامل ادامه دارد
ثابت کلیدهای حالت:
- جلوگیری از اشتباه تایپی در سیستم های چند عاملی
- به عنوان قرارداد بین نمایندگان عمل کنید
- زمانی که عامل ها داده ها را به اشتراک می گذارند بسیار مهم است
Async + Thread Pools:
-
async defبه ابزارها اجازه می دهد تا اجرا را متوقف کنند - Thread Pools کار متصل به CPU را در پس زمینه اجرا می کند
- آنها با هم حلقه رویداد را پاسخگو نگه می دارند
توابع کمکی:
- کمک های همگام سازی را از ابزارهای همگام جدا کنید
- کد را قابل آزمایش و استفاده مجدد می کند
دستورالعمل نماینده:
- دستورالعمل های دقیق از اشتباهات رایج LLM جلوگیری می کند
- صریح در مورد کارهایی که نباید انجام دهید (کد را اصلاح نکنید)
- مراحل گردش کار را برای ثبات پاک کنید
بعد چه است
در ماژول 5، شما اضافه خواهید کرد:
- جستجوگر سبک که کد را از حالت می خواند
- تست دونده که در واقع تست ها را اجرا می کند
- سینتی سایزر بازخورد که تمام تحلیل ها را ترکیب می کند
خواهید دید که چگونه حالت از طریق یک خط لوله متوالی جریان می یابد، و چرا الگوی ثابت ها زمانی که چندین عامل داده های مشابهی را می خوانند و می نویسند اهمیت دارد.
5. ساخت یک خط لوله: چندین عامل با هم کار می کنند
مقدمه
در ماژول 4، شما یک عامل واحد ساختید که ساختار کد را تجزیه و تحلیل می کند. اما بررسی جامع کد به چیزی بیش از تجزیه نیاز دارد - شما به بررسی سبک، اجرای آزمایش و ترکیب بازخورد هوشمند نیاز دارید.
این ماژول خط لوله ای از 4 عامل ایجاد می کند که به طور متوالی با هم کار می کنند و هر کدام به تجزیه و تحلیل تخصصی کمک می کنند:
- تحلیلگر کد (از ماژول 4) - ساختار را تجزیه می کند
- Style Checker - موارد نقض سبک را شناسایی می کند
- Test Runner - تست ها را اجرا و تایید می کند
- فیدبک سینت سایزر - همه چیز را به بازخورد عملی ترکیب می کند
مفهوم کلیدی: حالت به عنوان کانال ارتباطی. هر عاملی آنچه را که عوامل قبلی برای بیان نوشته اند می خواند، تجزیه و تحلیل خود را اضافه می کند و حالت غنی شده را به عامل بعدی منتقل می کند. الگوی ثابت ها از ماژول 4 زمانی حیاتی می شود که چندین عامل داده ها را به اشتراک بگذارند.
پیش‌نمایش آنچه می‌سازید: ارسال کد آشفته ← جریان وضعیت تماشا از طریق 4 عامل ← دریافت گزارش جامع با بازخورد شخصی‌شده بر اساس الگوهای گذشته.
مرحله 1: ابزار Style Checker + Agent را اضافه کنید
بررسیگر سبک، نقض PEP 8 را با استفاده از pycodestyle شناسایی می‌کند - یک ورقه قطعی، نه تفسیر مبتنی بر LLM.
ابزار بررسی سبک را اضافه کنید
👉 باز کنید
code_review_assistant/tools.py
👉 پیدا کنید:
# MODULE_5_STEP_1_STYLE_CHECKER_TOOL
👉 آن خط واحد را با:
async def check_code_style(code: str, tool_context: ToolContext) -> Dict[str, Any]:
"""
Checks code style compliance using pycodestyle (PEP 8).
Args:
code: Python source code to check (or will retrieve from state)
tool_context: ADK tool context
Returns:
Dictionary containing style score and issues
"""
logger.info("Tool: Checking code style...")
try:
# Retrieve code from state if not provided
if not code:
code = tool_context.state.get(StateKeys.CODE_TO_REVIEW, '')
if not code:
return {
"status": "error",
"message": "No code provided or found in state"
}
# Run style check in thread pool
loop = asyncio.get_event_loop()
with ThreadPoolExecutor() as executor:
result = await loop.run_in_executor(
executor, _perform_style_check, code
)
# Store results in state
tool_context.state[StateKeys.STYLE_SCORE] = result['score']
tool_context.state[StateKeys.STYLE_ISSUES] = result['issues']
tool_context.state[StateKeys.STYLE_ISSUE_COUNT] = result['issue_count']
logger.info(f"Tool: Style check complete - Score: {result['score']}/100, "
f"Issues: {result['issue_count']}")
return result
except Exception as e:
error_msg = f"Style check failed: {str(e)}"
logger.error(f"Tool: {error_msg}", exc_info=True)
# Set default values on error
tool_context.state[StateKeys.STYLE_SCORE] = 0
tool_context.state[StateKeys.STYLE_ISSUES] = []
return {
"status": "error",
"message": error_msg,
"score": 0
}
👉 حالا به انتهای فایل رفته و پیدا کنید:
# MODULE_5_STEP_1_STYLE_HELPERS
👉 آن خط واحد را با توابع کمکی جایگزین کنید:
def _perform_style_check(code: str) -> Dict[str, Any]:
"""Helper to perform style check in thread pool."""
import io
import sys
with tempfile.NamedTemporaryFile(mode='w', suffix='.py', delete=False) as tmp:
tmp.write(code)
tmp_path = tmp.name
try:
# Capture stdout to get pycodestyle output
old_stdout = sys.stdout
sys.stdout = captured_output = io.StringIO()
style_guide = pycodestyle.StyleGuide(
quiet=False, # We want output
max_line_length=100,
ignore=['E501', 'W503']
)
result = style_guide.check_files([tmp_path])
# Restore stdout
sys.stdout = old_stdout
# Parse captured output
output = captured_output.getvalue()
issues = []
for line in output.strip().split('\n'):
if line and ':' in line:
parts = line.split(':', 4)
if len(parts) >= 4:
try:
issues.append({
'line': int(parts[1]),
'column': int(parts[2]),
'code': parts[3].split()[0] if len(parts) > 3 else 'E000',
'message': parts[3].strip() if len(parts) > 3 else 'Unknown error'
})
except (ValueError, IndexError):
pass
# Add naming convention checks
try:
tree = ast.parse(code)
naming_issues = _check_naming_conventions(tree)
issues.extend(naming_issues)
except SyntaxError:
pass # Syntax errors will be caught elsewhere
# Calculate weighted score
score = _calculate_style_score(issues)
return {
"status": "success",
"score": score,
"issue_count": len(issues),
"issues": issues[:10], # First 10 issues
"summary": f"Style score: {score}/100 with {len(issues)} violations"
}
finally:
if os.path.exists(tmp_path):
os.unlink(tmp_path)
def _check_naming_conventions(tree: ast.AST) -> List[Dict[str, Any]]:
"""Check PEP 8 naming conventions."""
naming_issues = []
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef):
# Skip private/protected methods and __main__
if not node.name.startswith('_') and node.name != node.name.lower():
naming_issues.append({
'line': node.lineno,
'column': node.col_offset,
'code': 'N802',
'message': f"N802 function name '{node.name}' should be lowercase"
})
elif isinstance(node, ast.ClassDef):
# Check if class name follows CapWords convention
if not node.name[0].isupper() or '_' in node.name:
naming_issues.append({
'line': node.lineno,
'column': node.col_offset,
'code': 'N801',
'message': f"N801 class name '{node.name}' should use CapWords convention"
})
return naming_issues
def _calculate_style_score(issues: List[Dict[str, Any]]) -> int:
"""Calculate weighted style score based on violation severity."""
if not issues:
return 100
# Define weights by error type
weights = {
'E1': 10, # Indentation errors
'E2': 3, # Whitespace errors
'E3': 5, # Blank line errors
'E4': 8, # Import errors
'E5': 5, # Line length
'E7': 7, # Statement errors
'E9': 10, # Syntax errors
'W2': 2, # Whitespace warnings
'W3': 2, # Blank line warnings
'W5': 3, # Line break warnings
'N8': 7, # Naming conventions
}
total_deduction = 0
for issue in issues:
code_prefix = issue['code'][:2] if len(issue['code']) >= 2 else 'E2'
weight = weights.get(code_prefix, 3)
total_deduction += weight
# Cap at 100 points deduction
return max(0, 100 - min(total_deduction, 100))
Style Checker Agent را اضافه کنید
👉 باز کنید
code_review_assistant/sub_agents/review_pipeline/style_checker.py
👉 پیدا کنید:
# MODULE_5_STEP_1_INSTRUCTION_PROVIDER
👉 آن خط واحد را با:
async def style_checker_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects state variables."""
template = """You are a code style expert focused on PEP 8 compliance.
Your task:
1. Use the check_code_style tool to validate PEP 8 compliance
2. The tool will retrieve the ORIGINAL code from state automatically
3. Report violations exactly as found
4. Present the results clearly and confidently
CRITICAL:
- The tool checks the code EXACTLY as provided by the user
- Do not suggest the code was modified or fixed
- Report actual violations found in the original code
- If there are style issues, they should be reported honestly
Call the check_code_style tool with an empty string for the code parameter,
as the tool will retrieve the code from state automatically.
When presenting results based on what the tool returns:
- State the exact score from the tool results
- If score >= 90: "Excellent style compliance!"
- If score 70-89: "Good style with minor improvements needed"
- If score 50-69: "Style needs attention"
- If score < 50: "Significant style improvements needed"
List the specific violations found (the tool will provide these):
- Show line numbers, error codes, and messages
- Focus on the top 10 most important issues
Previous analysis: {structure_analysis_summary}
Format your response as:
## Style Analysis Results
- Style Score: [exact score]/100
- Total Issues: [count]
- Assessment: [your assessment based on score]
## Top Style Issues
[List issues with line numbers and descriptions]
## Recommendations
[Specific fixes for the most critical issues]"""
return await instructions_utils.inject_session_state(template, context)
👉 پیدا کنید:
# MODULE_5_STEP_1_STYLE_CHECKER_AGENT
👉 آن خط واحد را با:
style_checker_agent = Agent(
name="StyleChecker",
model=config.worker_model,
description="Checks Python code style against PEP 8 guidelines",
instruction=style_checker_instruction_provider,
tools=[FunctionTool(func=check_code_style)],
output_key="style_check_summary"
)
مرحله 2: عامل Test Runner را اضافه کنید
اجرای آزمایشی تست های جامعی را تولید می کند و آنها را با استفاده از مجری کد داخلی اجرا می کند.
👉 باز کنید
code_review_assistant/sub_agents/review_pipeline/test_runner.py
👉 پیدا کنید:
# MODULE_5_STEP_2_INSTRUCTION_PROVIDER
👉 آن خط واحد را با:
async def test_runner_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects the code_to_review directly."""
template = """You are a testing specialist who creates and runs tests for Python code.
THE CODE TO TEST IS:
{code_to_review}
YOUR TASK:
1. Understand what the function appears to do based on its name and structure
2. Generate comprehensive tests (15-20 test cases)
3. Execute the tests using your code executor
4. Analyze results to identify bugs vs expected behavior
5. Output a detailed JSON analysis
TESTING METHODOLOGY:
- Test with the most natural interpretation first
- When something fails, determine if it's a bug or unusual design
- Test edge cases, boundaries, and error scenarios
- Document any surprising behavior
Execute your tests and output ONLY valid JSON with this structure:
- "test_summary": object with "total_tests_run", "tests_passed", "tests_failed", "tests_with_errors", "critical_issues_found"
- "critical_issues": array of objects, each with "type", "description", "example_input", "expected_behavior", "actual_behavior", "severity"
- "test_categories": object with "basic_functionality", "edge_cases", "error_handling" (each containing "passed", "failed", "errors" counts)
- "function_behavior": object with "apparent_purpose", "actual_interface", "unexpected_requirements"
- "verdict": object with "status" (WORKING/BUGGY/BROKEN), "confidence" (high/medium/low), "recommendation"
Do NOT output the test code itself, only the JSON analysis."""
return await instructions_utils.inject_session_state(template, context)
👉 پیدا کنید:
# MODULE_5_STEP_2_TEST_RUNNER_AGENT
👉 آن خط واحد را با:
test_runner_agent = Agent(
name="TestRunner",
model=config.critic_model,
description="Generates and runs tests for Python code using safe code execution",
instruction=test_runner_instruction_provider,
code_executor=BuiltInCodeExecutor(),
output_key="test_execution_summary"
)
مرحله 3: درک حافظه برای یادگیری متقابل
قبل از ساختن سینتی سایزر بازخورد، باید تفاوت بین حالت و حافظه را درک کنید - دو مکانیسم ذخیره سازی مختلف برای دو هدف متفاوت.
حالت در مقابل حافظه: تمایز کلیدی
بیایید با یک مثال عینی از بررسی کد توضیح دهیم:
وضعیت (فقط جلسه فعلی):
# Data from THIS review session
tool_context.state[StateKeys.STYLE_ISSUES] = [
{"line": 5, "code": "E231", "message": "missing whitespace"},
{"line": 12, "code": "E701", "message": "multiple statements"}
]
- محدوده: فقط این مکالمه
- هدف: انتقال داده ها بین عوامل در خط لوله فعلی
- زندگی در: شی
Session - مادام العمر: با پایان جلسه کنار گذاشته می شود
حافظه (تمام جلسات گذشته):
# Learned from 50 previous reviews
"User frequently forgets docstrings on helper functions"
"User tends to write long functions (avg 45 lines)"
"User improved error handling after feedback in session #23"
- محدوده: تمام جلسات گذشته برای این کاربر
- هدف: یادگیری الگوها، ارائه بازخورد شخصی
- زندگی در:
MemoryService - طول عمر: در طول جلسات، قابل جستجو است
چرا بازخورد به هر دو مورد نیاز دارد:
تصور کنید سینت سایزر بازخورد ایجاد می کند:
فقط با استفاده از State (بررسی فعلی):
"Function `calculate_total` has no docstring."
بازخورد عمومی، مکانیکی.
استفاده از حالت + حافظه (الگوهای فعلی + گذشته):
"Function `calculate_total` has no docstring. This is the 4th review
where helper functions lacked documentation. Consider adding docstrings
as you write functions, not afterwards - you mentioned in our last
session that you find it easier that way."
شخصی، متنی، بهبود مراجع در طول زمان.
برای استقرار تولید، گزینه هایی دارید :
گزینه 1: VertexAiMemoryBankService (پیشرفته)
- کاری که انجام می دهد: استخراج حقایق معنادار از مکالمات به کمک LLM
- جستجو: جستجوی معنایی (معنا را درک می کند، نه فقط کلمات کلیدی)
- مدیریت حافظه: به طور خودکار حافظه ها را در طول زمان یکپارچه و به روز می کند
- مورد نیاز: راه اندازی Google Cloud Project + Agent Engine
- زمانی استفاده کنید: خاطرات پیچیده، در حال تکامل و شخصی می‌خواهید
- مثال: "کاربر برنامه نویسی کاربردی را ترجیح می دهد" (استخراج از 10 مکالمه در مورد سبک کد)
گزینه 2: با InMemoryMemoryService + Persistent Sessions ادامه دهید
- چه کاری انجام می دهد: تاریخچه مکالمه کامل را برای جستجوی کلمه کلیدی ذخیره می کند
- جستجو: تطبیق کلیدواژه اصلی در جلسات گذشته
- مدیریت حافظه: شما کنترل می کنید که چه چیزی ذخیره می شود (از طریق
add_session_to_memory) - مورد نیاز: فقط یک
SessionServiceدائمی (مانندVertexAiSessionServiceیاDatabaseSessionService) - استفاده از زمانی: به جستجوی ساده در میان مکالمات گذشته بدون پردازش LLM نیاز دارید
- مثال: جستجوی "docstring" تمام جلساتی را که آن کلمه را ذکر می‌کنند برمی‌گرداند
چگونه حافظه پر می شود
پس از تکمیل هر بازبینی کد:
# At the end of a session (typically in your application code)
await memory_service.add_session_to_memory(session)
چه اتفاقی می افتد:
- InMemoryMemoryService: رویدادهای جلسه کامل را برای جستجوی کلمه کلیدی ذخیره می کند
- VertexAiMemoryBankService: LLM حقایق کلیدی را استخراج می کند، با حافظه های موجود ادغام می شود
سپس جلسات آینده می توانند پرس و جو کنند:
# In a tool, search for relevant past feedback
results = tool_context.search_memory("feedback about docstrings")
مرحله 4: ابزار و عامل سنتز کننده بازخورد را اضافه کنید
سینت سایزر بازخورد پیچیده ترین عامل در خط لوله است. این سه ابزار را هماهنگ می کند، از دستورالعمل های پویا استفاده می کند و حالت، حافظه و مصنوعات را با هم ترکیب می کند.
سه ابزار سینتی سایزر را اضافه کنید
👉 باز کنید
code_review_assistant/tools.py
👉 پیدا کنید:
# MODULE_5_STEP_4_SEARCH_PAST_FEEDBACK
👉 جایگزینی با ابزار 1 - جستجوی حافظه (نسخه تولیدی):
async def search_past_feedback(developer_id: str, tool_context: ToolContext) -> Dict[str, Any]:
"""
Search for past feedback in memory service.
Args:
developer_id: ID of the developer (defaults to "default_user")
tool_context: ADK tool context with potential memory service access
Returns:
Dictionary containing feedback search results
"""
logger.info(f"Tool: Searching for past feedback for developer {developer_id}...")
try:
# Default developer ID if not provided
if not developer_id:
developer_id = tool_context.state.get(StateKeys.USER_ID, 'default_user')
# Check if memory service is available
if hasattr(tool_context, 'search_memory'):
try:
# Perform structured searches
queries = [
f"developer:{developer_id} code review feedback",
f"developer:{developer_id} common issues",
f"developer:{developer_id} improvements"
]
all_feedback = []
patterns = {
'common_issues': [],
'improvements': [],
'strengths': []
}
for query in queries:
search_result = await tool_context.search_memory(query)
if search_result and hasattr(search_result, 'memories'):
for memory in search_result.memories[:5]:
memory_text = memory.text if hasattr(memory, 'text') else str(memory)
all_feedback.append(memory_text)
# Extract patterns
if 'style' in memory_text.lower():
patterns['common_issues'].append('style compliance')
if 'improved' in memory_text.lower():
patterns['improvements'].append('showing improvement')
if 'excellent' in memory_text.lower():
patterns['strengths'].append('consistent quality')
# Store in state
tool_context.state[StateKeys.PAST_FEEDBACK] = all_feedback
tool_context.state[StateKeys.FEEDBACK_PATTERNS] = patterns
logger.info(f"Tool: Found {len(all_feedback)} past feedback items")
return {
"status": "success",
"feedback_found": True,
"count": len(all_feedback),
"summary": " | ".join(all_feedback[:3]) if all_feedback else "No feedback",
"patterns": patterns
}
except Exception as e:
logger.warning(f"Tool: Memory search error: {e}")
# Fallback: Check state for cached feedback
cached_feedback = tool_context.state.get(StateKeys.USER_PAST_FEEDBACK_CACHE, [])
if cached_feedback:
tool_context.state[StateKeys.PAST_FEEDBACK] = cached_feedback
return {
"status": "success",
"feedback_found": True,
"count": len(cached_feedback),
"summary": "Using cached feedback",
"patterns": {}
}
# No feedback found
tool_context.state[StateKeys.PAST_FEEDBACK] = []
logger.info("Tool: No past feedback found")
return {
"status": "success",
"feedback_found": False,
"message": "No past feedback available - this appears to be a first submission",
"patterns": {}
}
except Exception as e:
error_msg = f"Feedback search error: {str(e)}"
logger.error(f"Tool: {error_msg}", exc_info=True)
tool_context.state[StateKeys.PAST_FEEDBACK] = []
return {
"status": "error",
"message": error_msg,
"feedback_found": False
}
👉 پیدا کنید:
# MODULE_5_STEP_4_UPDATE_GRADING_PROGRESS
👉 جایگزینی با ابزار 2 - ردیاب درجه بندی (نسخه تولیدی):
async def update_grading_progress(tool_context: ToolContext) -> Dict[str, Any]:
"""
Updates grading progress counters and metrics in state.
"""
logger.info("Tool: Updating grading progress...")
try:
current_time = datetime.now().isoformat()
# Build all state changes
state_updates = {}
# Temporary (invocation-level) state
state_updates[StateKeys.TEMP_PROCESSING_TIMESTAMP] = current_time
# Session-level state
attempts = tool_context.state.get(StateKeys.GRADING_ATTEMPTS, 0) + 1
state_updates[StateKeys.GRADING_ATTEMPTS] = attempts
state_updates[StateKeys.LAST_GRADING_TIME] = current_time
# User-level persistent state
lifetime_submissions = tool_context.state.get(StateKeys.USER_TOTAL_SUBMISSIONS, 0) + 1
state_updates[StateKeys.USER_TOTAL_SUBMISSIONS] = lifetime_submissions
state_updates[StateKeys.USER_LAST_SUBMISSION_TIME] = current_time
# Calculate improvement metrics
current_style_score = tool_context.state.get(StateKeys.STYLE_SCORE, 0)
last_style_score = tool_context.state.get(StateKeys.USER_LAST_STYLE_SCORE, 0)
score_improvement = current_style_score - last_style_score
state_updates[StateKeys.USER_LAST_STYLE_SCORE] = current_style_score
state_updates[StateKeys.SCORE_IMPROVEMENT] = score_improvement
# Track test results if available
test_results = tool_context.state.get(StateKeys.TEST_EXECUTION_SUMMARY, {})
# Parse if it's a string
if isinstance(test_results, str):
try:
test_results = json.loads(test_results)
except:
test_results = {}
if test_results and test_results.get('test_summary', {}).get('total_tests_run', 0) > 0:
summary = test_results['test_summary']
total = summary.get('total_tests_run', 0)
passed = summary.get('tests_passed', 0)
if total > 0:
pass_rate = (passed / total) * 100
state_updates[StateKeys.USER_LAST_TEST_PASS_RATE] = pass_rate
# Apply all updates atomically
for key, value in state_updates.items():
tool_context.state[key] = value
logger.info(f"Tool: Progress updated - Attempt #{attempts}, "
f"Lifetime: {lifetime_submissions}")
return {
"status": "success",
"session_attempts": attempts,
"lifetime_submissions": lifetime_submissions,
"timestamp": current_time,
"improvement": {
"style_score_change": score_improvement,
"direction": "improved" if score_improvement > 0 else "declined"
},
"summary": f"Attempt #{attempts} recorded, {lifetime_submissions} total submissions"
}
except Exception as e:
error_msg = f"Progress update error: {str(e)}"
logger.error(f"Tool: {error_msg}", exc_info=True)
return {
"status": "error",
"message": error_msg
}
👉 پیدا کنید:
# MODULE_5_STEP_4_SAVE_GRADING_REPORT
👉 جایگزینی با ابزار 3 - مصنوع ساز (نسخه تولید):
async def save_grading_report(feedback_text: str, tool_context: ToolContext) -> Dict[str, Any]:
"""
Saves a detailed grading report as an artifact.
Args:
feedback_text: The feedback text to include in the report
tool_context: ADK tool context for state management
Returns:
Dictionary containing save status and details
"""
logger.info("Tool: Saving grading report...")
try:
# Gather all relevant data from state
code = tool_context.state.get(StateKeys.CODE_TO_REVIEW, '')
analysis = tool_context.state.get(StateKeys.CODE_ANALYSIS, {})
style_score = tool_context.state.get(StateKeys.STYLE_SCORE, 0)
style_issues = tool_context.state.get(StateKeys.STYLE_ISSUES, [])
# Get test results
test_results = tool_context.state.get(StateKeys.TEST_EXECUTION_SUMMARY, {})
# Parse if it's a string
if isinstance(test_results, str):
try:
test_results = json.loads(test_results)
except:
test_results = {}
timestamp = datetime.now().isoformat()
# Create comprehensive report dictionary
report = {
'timestamp': timestamp,
'grading_attempt': tool_context.state.get(StateKeys.GRADING_ATTEMPTS, 1),
'code': {
'content': code,
'line_count': len(code.splitlines()),
'hash': hashlib.md5(code.encode()).hexdigest()
},
'analysis': analysis,
'style': {
'score': style_score,
'issues': style_issues[:5] # First 5 issues
},
'tests': test_results,
'feedback': feedback_text,
'improvements': {
'score_change': tool_context.state.get(StateKeys.SCORE_IMPROVEMENT, 0),
'from_last_score': tool_context.state.get(StateKeys.USER_LAST_STYLE_SCORE, 0)
}
}
# Convert report to JSON string
report_json = json.dumps(report, indent=2)
report_part = types.Part.from_text(text=report_json)
# Try to save as artifact if the service is available
if hasattr(tool_context, 'save_artifact'):
try:
# Generate filename with timestamp (replace colons for filesystem compatibility)
filename = f"grading_report_{timestamp.replace(':', '-')}.json"
# Save the main report
version = await tool_context.save_artifact(filename, report_part)
# Also save a "latest" version for easy access
await tool_context.save_artifact("latest_grading_report.json", report_part)
logger.info(f"Tool: Report saved as {filename} (version {version})")
# Store report in state as well for redundancy
tool_context.state[StateKeys.USER_LAST_GRADING_REPORT] = report
return {
"status": "success",
"artifact_saved": True,
"filename": filename,
"version": str(version),
"size": len(report_json),
"summary": f"Report saved as {filename}"
}
except Exception as artifact_error:
logger.warning(f"Artifact service error: {artifact_error}, falling back to state storage")
# Continue to fallback below
# Fallback: Store in state if artifact service is not available or failed
tool_context.state[StateKeys.USER_LAST_GRADING_REPORT] = report
logger.info("Tool: Report saved to state (artifact service not available)")
return {
"status": "success",
"artifact_saved": False,
"message": "Report saved to state only",
"size": len(report_json),
"summary": "Report saved to session state"
}
except Exception as e:
error_msg = f"Report save error: {str(e)}"
logger.error(f"Tool: {error_msg}", exc_info=True)
# Still try to save minimal data to state
try:
tool_context.state[StateKeys.USER_LAST_GRADING_REPORT] = {
'error': error_msg,
'feedback': feedback_text,
'timestamp': datetime.now().isoformat()
}
except:
pass
return {
"status": "error",
"message": error_msg,
"artifact_saved": False,
"summary": f"Failed to save report: {error_msg}"
}
عامل Synthesizer را ایجاد کنید
👉 باز کنید
code_review_assistant/sub_agents/review_pipeline/feedback_synthesizer.py
👉 پیدا کنید:
# MODULE_5_STEP_4_INSTRUCTION_PROVIDER
👉 با ارائه دهنده دستورالعمل تولید جایگزین کنید:
async def feedback_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects state variables."""
template = """You are an expert code reviewer and mentor providing constructive, educational feedback.
CONTEXT FROM PREVIOUS AGENTS:
- Structure analysis summary: {structure_analysis_summary}
- Style check summary: {style_check_summary}
- Test execution summary: {test_execution_summary}
YOUR TASK requires these steps IN ORDER:
1. Call search_past_feedback tool with developer_id="default_user"
2. Call update_grading_progress tool with no parameters
3. Carefully analyze the test results to understand what really happened
4. Generate comprehensive feedback following the structure below
5. Call save_grading_report tool with the feedback_text parameter
6. Return the feedback as your final output
CRITICAL - Understanding Test Results:
The test_execution_summary contains structured JSON. Parse it carefully:
- tests_passed = Code worked correctly
- tests_failed = Code produced wrong output
- tests_with_errors = Code crashed
- critical_issues = Fundamental problems with the code
If critical_issues array contains items, these are serious bugs that need fixing.
Do NOT count discovering bugs as test successes.
FEEDBACK STRUCTURE TO FOLLOW:
## 📊 Summary
Provide an honest assessment. Be encouraging but truthful about problems found.
## ✅ Strengths
List 2-3 things done well, referencing specific code elements.
## 📈 Code Quality Analysis
### Structure & Organization
Comment on code organization, readability, and documentation.
### Style Compliance
Report the actual style score and any specific issues.
### Test Results
Report the actual test results accurately:
- If critical_issues exist, report them as bugs to fix
- Be clear: "X tests passed, Y critical issues were found"
- List each critical issue
- Don't hide or minimize problems
## 💡 Recommendations for Improvement
Based on the analysis, provide specific actionable fixes.
If critical issues exist, fixing them is top priority.
## 🎯 Next Steps
Prioritized action list based on severity of issues.
## 💬 Encouragement
End with encouragement while being honest about what needs fixing.
Remember: Complete ALL steps including calling save_grading_report."""
return await instructions_utils.inject_session_state(template, context)
👉 پیدا کنید:
# MODULE_5_STEP_4_SYNTHESIZER_AGENT
👉 جایگزین کنید با:
feedback_synthesizer_agent = Agent(
name="FeedbackSynthesizer",
model=config.critic_model,
description="Synthesizes all analysis into constructive, personalized feedback",
instruction=feedback_instruction_provider,
tools=[
FunctionTool(func=search_past_feedback),
FunctionTool(func=update_grading_progress),
FunctionTool(func=save_grading_report)
],
output_key="final_feedback"
)
مرحله 5: خط لوله را سیم کشی کنید
اکنون هر چهار عامل را به یک خط لوله متوالی متصل کرده و عامل ریشه را ایجاد کنید.
👉 باز کنید
code_review_assistant/agent.py
👉 واردات های لازم را در بالای فایل (بعد از واردات موجود) اضافه کنید:
from google.adk.agents import Agent, SequentialAgent
from code_review_assistant.sub_agents.review_pipeline.code_analyzer import code_analyzer_agent
from code_review_assistant.sub_agents.review_pipeline.style_checker import style_checker_agent
from code_review_assistant.sub_agents.review_pipeline.test_runner import test_runner_agent
from code_review_assistant.sub_agents.review_pipeline.feedback_synthesizer import feedback_synthesizer_agent
اکنون فایل شما باید به شکل زیر باشد:
"""
Main agent orchestration for the Code Review Assistant.
"""
from google.adk.agents import Agent, SequentialAgent
from .config import config
from code_review_assistant.sub_agents.review_pipeline.code_analyzer import code_analyzer_agent
from code_review_assistant.sub_agents.review_pipeline.style_checker import style_checker_agent
from code_review_assistant.sub_agents.review_pipeline.test_runner import test_runner_agent
from code_review_assistant.sub_agents.review_pipeline.feedback_synthesizer import feedback_synthesizer_agent
# MODULE_5_STEP_5_CREATE_PIPELINE
# MODULE_6_STEP_5_CREATE_FIX_LOOP
# MODULE_6_STEP_5_UPDATE_ROOT_AGENT
👉 پیدا کنید:
# MODULE_5_STEP_5_CREATE_PIPELINE
👉 آن خط واحد را با:
# Create sequential pipeline
code_review_pipeline = SequentialAgent(
name="CodeReviewPipeline",
description="Complete code review pipeline with analysis, testing, and feedback",
sub_agents=[
code_analyzer_agent,
style_checker_agent,
test_runner_agent,
feedback_synthesizer_agent
]
)
# Root agent - coordinates the review pipeline
root_agent = Agent(
name="CodeReviewAssistant",
model=config.worker_model,
description="An intelligent code review assistant that analyzes Python code and provides educational feedback",
instruction="""You are a specialized Python code review assistant focused on helping developers improve their code quality.
When a user provides Python code for review:
1. Immediately delegate to CodeReviewPipeline and pass the code EXACTLY as it was provided by the user.
2. The pipeline will handle all analysis and feedback
3. Return ONLY the final feedback from the pipeline - do not add any commentary
When a user asks what you can do or asks general questions:
- Explain your capabilities for code review
- Do NOT trigger the pipeline for non-code messages
The pipeline handles everything for code review - just pass through its final output.""",
sub_agents=[code_review_pipeline],
output_key="assistant_response"
)
مرحله 6: خط لوله کامل را آزمایش کنید
وقت آن است که ببینیم هر چهار نماینده با هم کار می کنند.
👉 سیستم را راه اندازی کنید:
adk web code_review_assistant
پس از اجرای دستور adk web ، باید خروجی را در ترمینال خود مشاهده کنید که نشان می دهد وب سرور ADK شروع شده است، مشابه این:
+-----------------------------------------------------------------------------+
| ADK Web Server started |
| |
| For local testing, access at http://localhost:8000. |
+-----------------------------------------------------------------------------+
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
👉 بعد، برای دسترسی به ADK Dev UI از مرورگر خود:
از نماد پیش‌نمایش وب (اغلب با یک فلش شبیه یک چشم یا مربع است) در نوار ابزار Cloud Shell (معمولاً بالا سمت راست)، Change port را انتخاب کنید. در پنجره پاپ آپ، پورت را روی 8000 تنظیم کنید و روی "Change and Preview" کلیک کنید. سپس Cloud Shell یک برگه یا پنجره جدید مرورگر را باز می کند که ADK Dev UI را نشان می دهد.
👉 نماینده اکنون در حال اجراست. ADK Dev UI در مرورگر شما ارتباط مستقیم شما با عامل است.
- هدف خود را انتخاب کنید: در منوی کشویی در بالای رابط کاربری، عامل
code_review_assistantرا انتخاب کنید.
👉 اعلان تست:
Please analyze the following:
def dfs_search_v1(graph, start, target):
"""Find if target is reachable from start."""
visited = set()
stack = start
while stack:
current = stack.pop()
if current == target:
return True
if current not in visited:
visited.add(current)
for neighbor in graph[current]:
if neighbor not in visited:
stack.append(neighbor)
return False
👉 خط لوله بررسی کد را در عمل مشاهده کنید:
وقتی تابع buggy dfs_search_v1 را ارسال می کنید، فقط یک پاسخ دریافت نمی کنید. شما شاهد خط لوله چند عاملی خود در کار هستید. خروجی جریانی که مشاهده می‌کنید، نتیجه چهار عامل تخصصی است که به ترتیب اجرا می‌شوند، هر کدام بر روی آخرین ساختمان.
در اینجا به تفکیک آنچه که هر عامل در بررسی نهایی و جامع کمک می‌کند، آورده شده است، و داده‌های خام را به اطلاعاتی قابل اجرا تبدیل می‌کند.
1. گزارش ساختاری کد تحلیلگر
ابتدا عامل CodeAnalyzer کد خام را دریافت می کند. نمی توان حدس زد که کد چه کاری انجام می دهد. از ابزار analyze_code_structure برای انجام تجزیه قطعی درخت نحو انتزاعی (AST) استفاده می کند.
خروجی آن داده های واقعی و خالص در مورد ساختار کد است:
The analysis of the provided code reveals the following:
Summary:
- Functions Found: 1
- Classes Found: 0
Key Structural Observations:
- A single function, dfs_search_v1, is defined.
- It includes a docstring: "Find if target is reachable from start."
- No syntax errors were detected.
Overall Code Organization Assessment:
- The code snippet is a well-defined, self-contained function.
⭐ ارزش: این مرحله اولیه یک پایه تمیز و قابل اعتماد برای سایر عوامل فراهم می کند. تأیید می کند که کد پایتون معتبر است و مؤلفه های دقیقی را که باید بررسی شوند را شناسایی می کند.
2. ممیزی PEP 8 The Style Checker's
بعد، عامل StyleChecker مسئولیت را بر عهده می گیرد. کد را از حالت اشتراک‌گذاری شده می‌خواند و از ابزار check_code_style استفاده می‌کند که از pycodestyle linter استفاده می‌کند.
خروجی آن یک امتیاز کیفی قابل سنجش و تخلفات خاص است:
Style Analysis Results
- Style Score: 88/100
- Total Issues: 6
- Assessment: Good style with minor improvements needed
Top Style Issues
- Line 5, W293: blank line contains whitespace
- Line 19, W292: no newline at end of file
⭐ ارزش: این عامل بازخورد عینی و غیرقابل مذاکره بر اساس استانداردهای جامعه (PEP 8) ارائه می دهد. سیستم امتیازدهی وزنی بلافاصله شدت مسائل را به کاربر می گوید.
3. کشف اشکال حیاتی The Test Runner's
اینجاست که سیستم از تحلیل سطحی فراتر می رود. عامل TestRunner مجموعه جامعی از آزمایش‌ها را برای اعتبارسنجی رفتار کد تولید و اجرا می‌کند.
خروجی آن یک شی JSON ساختار یافته است که حاوی یک حکم نکوهشی است:
{
"critical_issues": [
{
"type": "Critical Bug",
"description": "The function's initialization `stack = start` is incorrect... When a common input like a string... is provided... the function crashes with an AttributeError.",
"severity": "Critical"
}
],
"verdict": {
"status": "BROKEN",
"confidence": "high",
"recommendation": "The function is fundamentally broken... the stack initialization line `stack = start` must be changed to `stack = [start]`."
}
}
⭐ ارزش: این مهم ترین بینش است. نماینده فقط حدس نمی زد. ثابت کرد که کد با اجرای آن شکسته شده است. این یک باگ ظریف اما حیاتی در زمان اجرا را کشف کرد که بازبین انسانی ممکن است به راحتی آن را از دست بدهد و علت دقیق و رفع مورد نیاز را مشخص کند.
4. گزارش نهایی فیدبک سینت سایزر
در نهایت، عامل FeedbackSynthesizer به عنوان هادی عمل می کند. این داده های ساختار یافته را از سه عامل قبلی می گیرد و یک گزارش واحد و کاربرپسند ایجاد می کند که هم تحلیلی و هم دلگرم کننده است.
خروجی آن بررسی نهایی و صیقلی است که می بینید:
📊 Summary
Great effort on implementing the Depth-First Search algorithm! ... However, a critical bug in the initialization of the stack prevents the function from working correctly...
✅ Strengths
- Good Algorithm Structure
- Correct Use of `visited` Set
📈 Code Quality Analysis
...
### Style Compliance
The style analysis returned a good score of 88/100.
...
### Test Results
The automated testing revealed a critical issue... The line `stack = start` directly assigns the input... which results in an `AttributeError`.
💡 Recommendations for Improvement
**Fix the Critical Stack Initialization Bug:**
- Incorrect Code: `stack = start`
- Correct Code: `stack = [start]`
💬 Encouragement
You are very close to a perfect implementation! The core logic of your DFS algorithm is sound, which is the hardest part.
⭐ ارزش: این عامل داده های فنی را به یک تجربه مفید و آموزشی تبدیل می کند. مهم ترین مسئله (اشکال) را اولویت بندی می کند، آن را به وضوح توضیح می دهد، راه حل دقیق را ارائه می دهد، و این کار را با لحنی دلگرم کننده انجام می دهد. این با موفقیت یافته های تمام مراحل قبلی را در یک کل منسجم و ارزشمند ادغام می کند.
این فرآیند چند مرحله ای قدرت یک خط لوله عامل را نشان می دهد. Instead of a single, monolithic response, you get a layered analysis where each agent performs a specialized, verifiable task. This leads to a review that is not only insightful but also deterministic, reliable, and deeply educational.
👉💻 Once you're done testing, return to your Cloud Shell Editor terminal and press Ctrl+C to stop the ADK Dev UI.
What You've Built
You now have a complete code review pipeline that:
✅ Parses code structure - deterministic AST analysis with helper functions
✅ Checks style - weighted scoring with naming conventions
✅ Runs tests - comprehensive test generation with structured JSON output
✅ Synthesizes feedback - integrates state + memory + artifacts
✅ Tracks progress - multi-tier state across invocations/sessions/users
✅ Learns over time - memory service for cross-session patterns
✅ Provides artifacts - downloadable JSON reports with complete audit trail
Key Concepts Mastered
Sequential Pipelines:
- Four agents executing in strict order
- Each enriches state for the next
- Dependencies determine execution sequence
Production Patterns:
- Helper function separation (sync in thread pools)
- Graceful degradation (fallback strategies)
- Multi-tier state management (temp/session/user)
- Dynamic instruction providers (context-aware)
- Dual storage (artifacts + state redundancy)
State as Communication:
- Constants prevent typos across agents
-
output_keywrites agent summaries to state - Later agents read via StateKeys
- State flows linearly through pipeline
Memory vs State:
- State: current session data
- Memory: patterns across sessions
- Different purposes, different lifetimes
Tool Orchestration:
- Single-tool agents (analyzer, style_checker)
- Built-in executors (test_runner)
- Multi-tool coordination (synthesizer)
Model Selection Strategy:
- Worker model: mechanical tasks (parsing, linting, routing)
- Critic model: reasoning tasks (testing, synthesis)
- Cost optimization through appropriate selection
بعد چه است
In Module 6, you'll build the fix pipeline :
- LoopAgent architecture for iterative fixing
- Exit conditions via escalation
- State accumulation across iterations
- Validation and retry logic
- Integration with review pipeline to offer fixes
You'll see how the same state patterns scale to complex iterative workflows where agents attempt multiple times until success, and how to coordinate multiple pipelines in a single application.
6. Adding the Fix Pipeline: Loop Architecture
مقدمه
In Module 5, you built a sequential review pipeline that analyzes code and provides feedback. But identifying problems is only half the solution - developers need help fixing them.
This module builds an automated fix pipeline that:
- Generates fixes based on review results
- Validates fixes by running comprehensive tests
- Retries automatically if fixes don't work (up to 3 attempts)
- Reports results with before/after comparisons
Key concept: LoopAgent for automatic retry. Unlike sequential agents that run once, a LoopAgent repeats its sub-agents until an exit condition is met or max iterations reached. Tools signal success by setting tool_context.actions.escalate = True .
Preview of what you'll build: Submit buggy code → review identifies issues → fix loop generates corrections → tests validate → retries if needed → final comprehensive report.
Core Concepts: LoopAgent vs Sequential
Sequential Pipeline (Module 5):
SequentialAgent(agents=[A, B, C])
# Executes: A → B → C → Done
- One-way flow
- Each agent runs exactly once
- No retry logic
Loop Pipeline (Module 6):
LoopAgent(agents=[A, B, C], max_iterations=3)
# Executes: A → B → C → (check exit) → A → B → C → (check exit) → ...
- Cyclic flow
- Agents can run multiple times
- Exits when:
- A tool sets
tool_context.actions.escalate = True(success) -
max_iterationsreached (safety limit) - Unhandled exception occurs (error)
- A tool sets
Why loops for code fixing:
Code fixes often need multiple attempts:
- First attempt : Fix obvious bugs (wrong variable types)
- Second attempt : Fix secondary issues revealed by tests (edge cases)
- Third attempt : Fine-tune and validate all tests pass
Without a loop, you'd need complex conditional logic in agent instructions. With LoopAgent , retry is automatic.
Architecture comparison:
Sequential (Module 5):
User → Review Pipeline → Feedback → Done
Loop (Module 6):
User → Review Pipeline → Feedback → Fix Pipeline
↓
┌──────────────┴──────────────┐
│ Fix Attempt Loop (1-3x) │
│ ┌─────────────────────┐ │
│ │ 1. Generate Fixes │ │
│ │ 2. Test Fixes │ │
│ │ 3. Validate & Exit? │────┼─→ If escalate=True
│ └─────────────────────┘ │ exit loop
│ ↓ If not │
│ Try Again (max 3) │
└─────────────────────────────┘
↓
4. Synthesize Final Report → Done
Step 1: Add Code Fixer Agent
The code fixer generates corrected Python code based on review results.
👉 Open
code_review_assistant/sub_agents/fix_pipeline/code_fixer.py
👉 Find:
# MODULE_6_STEP_1_CODE_FIXER_INSTRUCTION_PROVIDER
👉 Replace that single line with:
async def code_fixer_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects state variables."""
template = """You are an expert code fixing specialist.
Original Code:
{code_to_review}
Analysis Results:
- Style Score: {style_score}/100
- Style Issues: {style_issues}
- Test Results: {test_execution_summary}
Based on the test results, identify and fix ALL issues including:
- Interface bugs (e.g., if start parameter expects wrong type)
- Logic errors (e.g., KeyError when accessing graph nodes)
- Style violations
- Missing documentation
YOUR TASK:
Generate the complete fixed Python code that addresses all identified issues.
CRITICAL INSTRUCTIONS:
- Output ONLY the corrected Python code
- Do NOT include markdown code blocks (```python)
- Do NOT include any explanations or commentary
- The output should be valid, executable Python code and nothing else
Common fixes to apply based on test results:
- If tests show AttributeError with 'pop', fix: stack = [start] instead of stack = start
- If tests show KeyError accessing graph, fix: use graph.get(current, [])
- Add docstrings if missing
- Fix any style violations identified
Output the complete fixed code now:"""
return await instructions_utils.inject_session_state(template, context)
👉 Find:
# MODULE_6_STEP_1_CODE_FIXER_AGENT
👉 Replace that single line with:
code_fixer_agent = Agent(
name="CodeFixer",
model=config.worker_model,
description="Generates comprehensive fixes for all identified code issues",
instruction=code_fixer_instruction_provider,
code_executor=BuiltInCodeExecutor(),
output_key="code_fixes"
)
Step 2: Add Fix Test Runner Agent
The fix test runner validates corrections by executing comprehensive tests on the fixed code.
👉 Open
code_review_assistant/sub_agents/fix_pipeline/fix_test_runner.py
👉 Find:
# MODULE_6_STEP_2_FIX_TEST_RUNNER_INSTRUCTION_PROVIDER
👉 Replace that single line with:
async def fix_test_runner_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that uses the clean code from the previous step."""
template = """You are responsible for validating the fixed code by running tests.
THE FIXED CODE TO TEST:
{code_fixes}
ORIGINAL TEST RESULTS: {test_execution_summary}
YOUR TASK:
1. Understand the fixes that were applied
2. Generate the same comprehensive tests (15-20 test cases)
3. Execute the tests on the FIXED code using your code executor
4. Compare results with original test results
5. Output a detailed JSON analysis
TESTING METHODOLOGY:
- Run the same tests that revealed issues in the original code
- Verify that previously failing tests now pass
- Ensure no regressions were introduced
- Document the improvement
Execute your tests and output ONLY valid JSON with this structure:
- "passed": number of tests that passed
- "failed": number of tests that failed
- "total": total number of tests
- "pass_rate": percentage as a number
- "comparison": object with "original_pass_rate", "new_pass_rate", "improvement"
- "newly_passing_tests": array of test names that now pass
- "still_failing_tests": array of test names still failing
Do NOT output the test code itself, only the JSON analysis."""
return await instructions_utils.inject_session_state(template, context)
👉 Find:
# MODULE_6_STEP_2_FIX_TEST_RUNNER_AGENT
👉 Replace that single line with:
fix_test_runner_agent = Agent(
name="FixTestRunner",
model=config.critic_model,
description="Runs comprehensive tests on fixed code to verify all issues are resolved",
instruction=fix_test_runner_instruction_provider,
code_executor=BuiltInCodeExecutor(),
output_key="fix_test_execution_summary"
)
Step 3: Add Fix Validator Agent
The validator checks if fixes were successful and decides whether to exit the loop.
Understanding the Tools
First, add the three tools the validator needs.
👉 Open
code_review_assistant/tools.py
👉 Find:
# MODULE_6_STEP_3_VALIDATE_FIXED_STYLE
👉 Replace with Tool 1 - Style Validator:
async def validate_fixed_style(tool_context: ToolContext) -> Dict[str, Any]:
"""
Validates style compliance of the fixed code.
Args:
tool_context: ADK tool context containing fixed code in state
Returns:
Dictionary with style validation results
"""
logger.info("Tool: Validating style of fixed code...")
try:
# Get the fixed code from state
code_fixes = tool_context.state.get(StateKeys.CODE_FIXES, '')
# Try to extract from markdown if present
if '```python' in code_fixes:
start = code_fixes.rfind('```python') + 9
end = code_fixes.rfind('```')
if start < end:
code_fixes = code_fixes[start:end].strip()
if not code_fixes:
return {
"status": "error",
"message": "No fixed code found in state"
}
# Store the extracted fixed code
tool_context.state[StateKeys.CODE_FIXES] = code_fixes
# Run style check on fixed code
loop = asyncio.get_event_loop()
with ThreadPoolExecutor() as executor:
style_result = await loop.run_in_executor(
executor, _perform_style_check, code_fixes
)
# Compare with original
original_score = tool_context.state.get(StateKeys.STYLE_SCORE, 0)
improvement = style_result['score'] - original_score
# Store results
tool_context.state[StateKeys.FIXED_STYLE_SCORE] = style_result['score']
tool_context.state[StateKeys.FIXED_STYLE_ISSUES] = style_result['issues']
logger.info(f"Tool: Fixed code style score: {style_result['score']}/100 "
f"(improvement: +{improvement})")
return {
"status": "success",
"fixed_style_score": style_result['score'],
"original_style_score": original_score,
"improvement": improvement,
"remaining_issues": style_result['issues'],
"perfect_style": style_result['score'] == 100
}
except Exception as e:
logger.error(f"Tool: Style validation failed: {e}", exc_info=True)
return {
"status": "error",
"message": str(e)
}
👉 Find:
# MODULE_6_STEP_3_COMPILE_FIX_REPORT
👉 Replace with Tool 2 - Report Compiler:
async def compile_fix_report(tool_context: ToolContext) -> Dict[str, Any]:
"""
Compiles comprehensive report of the fix process.
Args:
tool_context: ADK tool context with all fix pipeline data
Returns:
Comprehensive fix report
"""
logger.info("Tool: Compiling comprehensive fix report...")
try:
# Gather all data
original_code = tool_context.state.get(StateKeys.CODE_TO_REVIEW, '')
code_fixes = tool_context.state.get(StateKeys.CODE_FIXES, '')
# Test results
original_tests = tool_context.state.get(StateKeys.TEST_EXECUTION_SUMMARY, {})
fixed_tests = tool_context.state.get(StateKeys.FIX_TEST_EXECUTION_SUMMARY, {})
# Parse if strings
if isinstance(original_tests, str):
try:
original_tests = json.loads(original_tests)
except:
original_tests = {}
if isinstance(fixed_tests, str):
try:
fixed_tests = json.loads(fixed_tests)
except:
fixed_tests = {}
# Extract pass rates
original_pass_rate = 0
if original_tests:
if 'pass_rate' in original_tests:
original_pass_rate = original_tests['pass_rate']
elif 'test_summary' in original_tests:
# Handle test_runner_agent's JSON structure
summary = original_tests['test_summary']
total = summary.get('total_tests_run', 0)
passed = summary.get('tests_passed', 0)
if total > 0:
original_pass_rate = (passed / total) * 100
elif 'passed' in original_tests and 'total' in original_tests:
if original_tests['total'] > 0:
original_pass_rate = (original_tests['passed'] / original_tests['total']) * 100
fixed_pass_rate = 0
all_tests_pass = False
if fixed_tests:
if 'pass_rate' in fixed_tests:
fixed_pass_rate = fixed_tests['pass_rate']
all_tests_pass = fixed_tests.get('failed', 1) == 0
elif 'passed' in fixed_tests and 'total' in fixed_tests:
if fixed_tests['total'] > 0:
fixed_pass_rate = (fixed_tests['passed'] / fixed_tests['total']) * 100
all_tests_pass = fixed_tests.get('failed', 0) == 0
# Style scores
original_style = tool_context.state.get(StateKeys.STYLE_SCORE, 0)
fixed_style = tool_context.state.get(StateKeys.FIXED_STYLE_SCORE, 0)
# Calculate improvements
test_improvement = {
'original_pass_rate': original_pass_rate,
'fixed_pass_rate': fixed_pass_rate,
'improvement': fixed_pass_rate - original_pass_rate,
'all_tests_pass': all_tests_pass
}
style_improvement = {
'original_score': original_style,
'fixed_score': fixed_style,
'improvement': fixed_style - original_style,
'perfect_style': fixed_style == 100
}
# Determine overall status
if all_tests_pass and style_improvement['perfect_style']:
fix_status = 'SUCCESSFUL'
status_emoji = '✅'
elif test_improvement['improvement'] > 0 or style_improvement['improvement'] > 0:
fix_status = 'PARTIAL'
status_emoji = '⚠️'
else:
fix_status = 'FAILED'
status_emoji = '❌'
# Build comprehensive report
report = {
'status': fix_status,
'status_emoji': status_emoji,
'timestamp': datetime.now().isoformat(),
'original_code': original_code,
'code_fixes': code_fixes,
'improvements': {
'tests': test_improvement,
'style': style_improvement
},
'summary': f"{status_emoji} Fix Status: {fix_status}\n"
f"Tests: {original_pass_rate:.1f}% → {fixed_pass_rate:.1f}%\n"
f"Style: {original_style}/100 → {fixed_style}/100"
}
# Store report in state
tool_context.state[StateKeys.FIX_REPORT] = report
tool_context.state[StateKeys.FIX_STATUS] = fix_status
logger.info(f"Tool: Fix report compiled - Status: {fix_status}")
logger.info(f"Tool: Test improvement: {original_pass_rate:.1f}% → {fixed_pass_rate:.1f}%")
logger.info(f"Tool: Style improvement: {original_style} → {fixed_style}")
return {
"status": "success",
"fix_status": fix_status,
"report": report
}
except Exception as e:
logger.error(f"Tool: Failed to compile fix report: {e}", exc_info=True)
return {
"status": "error",
"message": str(e)
}
👉 Find:
# MODULE_6_STEP_3_EXIT_FIX_LOOP
👉 Replace with Tool 3 - Loop Exit Signal:
def exit_fix_loop(tool_context: ToolContext) -> Dict[str, Any]:
"""
Signal that fixing is complete and should exit the loop.
Args:
tool_context: ADK tool context
Returns:
Confirmation message
"""
logger.info("Tool: Setting escalate flag to exit fix loop")
# This is the critical line that exits the LoopAgent
tool_context.actions.escalate = True
return {
"status": "success",
"message": "Fix complete, exiting loop"
}
Create the Validator Agent
👉 Open
code_review_assistant/sub_agents/fix_pipeline/fix_validator.py
👉 Find:
# MODULE_6_STEP_3_FIX_VALIDATOR_INSTRUCTION_PROVIDER
👉 Replace that single line with:
async def fix_validator_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects state variables."""
template = """You are the final validation specialist for code fixes.
You have access to:
- Original issues from initial review
- Applied fixes: {code_fixes}
- Test results after fix: {fix_test_execution_summary}
- All state data from the fix process
Your responsibilities:
1. Use validate_fixed_style tool to check style compliance of fixed code
- Pass no arguments, it will retrieve fixed code from state
2. Use compile_fix_report tool to generate comprehensive report
- Pass no arguments, it will gather all data from state
3. Based on the report, determine overall fix status:
- ✅ SUCCESSFUL: All tests pass, style score 100
- ⚠️ PARTIAL: Improvements made but issues remain
- ❌ FAILED: Fix didn't work or made things worse
4. CRITICAL: If status is SUCCESSFUL, call the exit_fix_loop tool to stop iterations
- This prevents unnecessary additional fix attempts
- If not successful, the loop will continue for another attempt
5. Provide clear summary of:
- What was fixed
- What improvements were achieved
- Any remaining issues requiring manual attention
Be precise and quantitative in your assessment.
"""
return await instructions_utils.inject_session_state(template, context)
👉 Find:
# MODULE_6_STEP_3_FIX_VALIDATOR_AGENT
👉 Replace that single line with:
fix_validator_agent = Agent(
name="FixValidator",
model=config.worker_model,
description="Validates fixes and generates final fix report",
instruction=fix_validator_instruction_provider,
tools=[
FunctionTool(func=validate_fixed_style),
FunctionTool(func=compile_fix_report),
FunctionTool(func=exit_fix_loop)
],
output_key="final_fix_report"
)
Step 4: Understanding LoopAgent Exit Conditions
The LoopAgent has three ways to exit:
1. Success Exit (via escalate)
# Inside any tool in the loop:
tool_context.actions.escalate = True
# Effect: Loop completes current iteration, then exits
# Use when: Fix is successful and no more attempts needed
Example flow:
Iteration 1:
CodeFixer → generates fixes
FixTestRunner → tests show 90% pass rate
FixValidator → compiles report, sees PARTIAL status
→ Does NOT set escalate
→ Loop continues
Iteration 2:
CodeFixer → refines fixes based on failures
FixTestRunner → tests show 100% pass rate
FixValidator → compiles report, sees SUCCESSFUL status
→ Calls exit_fix_loop() which sets escalate = True
→ Loop exits after this iteration
2. Max Iterations Exit
LoopAgent(
name="FixAttemptLoop",
sub_agents=[...],
max_iterations=3 # Safety limit
)
# Effect: After 3 complete iterations, loop exits regardless of escalate
# Use when: Prevent infinite loops if fixes never succeed
Example flow:
Iteration 1: PARTIAL (continue)
Iteration 2: PARTIAL (continue)
Iteration 3: PARTIAL (but max reached)
→ Loop exits, synthesizer presents best attempt
3. Error Exit
# If any agent throws unhandled exception:
raise Exception("Unexpected error")
# Effect: Loop exits immediately with error state
# Use when: Critical failure that can't be recovered
State Evolution Across Iterations:
Each iteration sees updated state from the previous attempt:
# Before Iteration 1:
state = {
"code_to_review": "def add(a,b):return a+b", # Original
"style_score": 40,
"test_execution_summary": {...}
}
# After Iteration 1:
state = {
"code_to_review": "def add(a,b):return a+b", # Unchanged
"code_fixes": "def add(a, b):\n return a + b", # NEW
"style_score": 40, # Unchanged
"fixed_style_score": 100, # NEW
"test_execution_summary": {...}, # Unchanged
"fix_test_execution_summary": {...} # NEW
}
# Iteration 2 starts with all this state
# If fixes still not perfect, code_fixes gets overwritten
چرا
escalate
Instead of Return Values:
# Bad: Using return value to signal exit
def validator_agent():
report = compile_report()
if report['status'] == 'SUCCESSFUL':
return {"exit": True} # How does loop know?
# Good: Using escalate
def validator_tool(tool_context):
report = compile_report()
if report['status'] == 'SUCCESSFUL':
tool_context.actions.escalate = True # Loop knows immediately
return {"report": report}
مزایا:
- Works from any tool, not just the last one
- Doesn't interfere with return data
- Clear semantic meaning
- Framework handles the exit logic
مرحله 5: خط لوله Fix را سیم کشی کنید
👉 Open
code_review_assistant/agent.py
👉 Add the fix pipeline imports (after the existing imports):
from google.adk.agents import LoopAgent # Add this to the existing Agent, SequentialAgent line
from code_review_assistant.sub_agents.fix_pipeline.code_fixer import code_fixer_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_test_runner import fix_test_runner_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_validator import fix_validator_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_synthesizer import fix_synthesizer_agent
Your imports should now be:
from google.adk.agents import Agent, SequentialAgent, LoopAgent
from .config import config
# Review pipeline imports (from Module 5)
from code_review_assistant.sub_agents.review_pipeline.code_analyzer import code_analyzer_agent
from code_review_assistant.sub_agents.review_pipeline.style_checker import style_checker_agent
from code_review_assistant.sub_agents.review_pipeline.test_runner import test_runner_agent
from code_review_assistant.sub_agents.review_pipeline.feedback_synthesizer import feedback_synthesizer_agent
# Fix pipeline imports (NEW)
from code_review_assistant.sub_agents.fix_pipeline.code_fixer import code_fixer_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_test_runner import fix_test_runner_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_validator import fix_validator_agent
from code_review_assistant.sub_agents.fix_pipeline.fix_synthesizer import fix_synthesizer_agent
👉 Find:
# MODULE_6_STEP_5_CREATE_FIX_LOOP
👉 Replace that single line with:
# Create the fix attempt loop (retries up to 3 times)
fix_attempt_loop = LoopAgent(
name="FixAttemptLoop",
sub_agents=[
code_fixer_agent, # Step 1: Generate fixes
fix_test_runner_agent, # Step 2: Validate with tests
fix_validator_agent # Step 3: Check success & possibly exit
],
max_iterations=3 # Try up to 3 times
)
# Wrap loop with synthesizer for final report
code_fix_pipeline = SequentialAgent(
name="CodeFixPipeline",
description="Automated code fixing pipeline with iterative validation",
sub_agents=[
fix_attempt_loop, # Try to fix (1-3 times)
fix_synthesizer_agent # Present final results (always runs once)
]
)
👉 Remove the existing
root_agent
تعریف:
root_agent = Agent(...)
👉 Find:
# MODULE_6_STEP_5_UPDATE_ROOT_AGENT
👉 Replace that single line with:
# Update root agent to include both pipelines
root_agent = Agent(
name="CodeReviewAssistant",
model=config.worker_model,
description="An intelligent code review assistant that analyzes Python code and provides educational feedback",
instruction="""You are a specialized Python code review assistant focused on helping developers improve their code quality.
When a user provides Python code for review:
1. Immediately delegate to CodeReviewPipeline and pass the code EXACTLY as it was provided by the user.
2. The pipeline will handle all analysis and feedback
3. Return ONLY the final feedback from the pipeline - do not add any commentary
After completing a review, if significant issues were identified:
- If style score < 100 OR tests are failing OR critical issues exist:
* Add at the end: "\n\n💡 I can fix these issues for you. Would you like me to do that?"
- If the user responds yes or requests fixes:
* Delegate to CodeFixPipeline
* Return the fix pipeline's complete output AS-IS
When a user asks what you can do or general questions:
- Explain your capabilities for code review and fixing
- Do NOT trigger the pipeline for non-code messages
The pipelines handle everything for code review and fixing - just pass through their final output.""",
sub_agents=[code_review_pipeline, code_fix_pipeline],
output_key="assistant_response"
)
مرحله 6: Fix Synthesizer Agent را اضافه کنید
The synthesizer creates a user-friendly presentation of fix results after the loop completes.
👉 Open
code_review_assistant/sub_agents/fix_pipeline/fix_synthesizer.py
👉 Find:
# MODULE_6_STEP_6_FIX_SYNTHESIZER_INSTRUCTION_PROVIDER
👉 Replace that single line with:
async def fix_synthesizer_instruction_provider(context: ReadonlyContext) -> str:
"""Dynamic instruction provider that injects state variables."""
template = """You are responsible for presenting the fix results to the user.
Based on the validation report: {final_fix_report}
Fixed code from state: {code_fixes}
Fix status: {fix_status}
Create a comprehensive yet friendly response that includes:
## 🔧 Fix Summary
[Overall status and key improvements - be specific about what was achieved]
## 📊 Metrics
- Test Results: [original pass rate]% → [new pass rate]%
- Style Score: [original]/100 → [new]/100
- Issues Fixed: X of Y
## ✅ What Was Fixed
[List each fixed issue with brief explanation of the correction made]
## 📝 Complete Fixed Code
[Include the complete, corrected code from state - this is critical]
## 💡 Explanation of Key Changes
[Brief explanation of the most important changes made and why]
[If any issues remain]
## ⚠️ Remaining Issues
[List what still needs manual attention]
## 🎯 Next Steps
[Guidance on what to do next - either use the fixed code or address remaining issues]
Save the fix report using save_fix_report tool before presenting.
Call it with no parameters - it will retrieve the report from state automatically.
Be encouraging about improvements while being honest about any remaining issues.
Focus on the educational aspect - help the user understand what was wrong and how it was fixed.
"""
return await instructions_utils.inject_session_state(template, context)
👉 Find:
# MODULE_6_STEP_6_FIX_SYNTHESIZER_AGENT
👉 Replace that single line with:
fix_synthesizer_agent = Agent(
name="FixSynthesizer",
model=config.critic_model,
description="Creates comprehensive user-friendly fix report",
instruction=fix_synthesizer_instruction_provider,
tools=[FunctionTool(func=save_fix_report)],
output_key="fix_summary"
)
👉 Add
save_fix_report
tool to
tools.py
:
👉 Find:
# MODULE_6_STEP_6_SAVE_FIX_REPORT
👉 Replace with:
async def save_fix_report(tool_context: ToolContext) -> Dict[str, Any]:
"""
Saves the fix report as an artifact.
Args:
tool_context: ADK tool context
Returns:
Save status
"""
logger.info("Tool: Saving fix report...")
try:
# Get the report from state
fix_report = tool_context.state.get(StateKeys.FIX_REPORT, {})
if not fix_report:
return {
"status": "error",
"message": "No fix report found in state"
}
# Convert to JSON
report_json = json.dumps(fix_report, indent=2)
report_part = types.Part.from_text(text=report_json)
# Generate filename
timestamp = datetime.now().isoformat().replace(':', '-')
filename = f"fix_report_{timestamp}.json"
# Try to save as artifact
if hasattr(tool_context, 'save_artifact'):
try:
version = await tool_context.save_artifact(filename, report_part)
await tool_context.save_artifact("latest_fix_report.json", report_part)
logger.info(f"Tool: Fix report saved as {filename}")
return {
"status": "success",
"filename": filename,
"version": str(version),
"size": len(report_json)
}
except Exception as e:
logger.warning(f"Could not save as artifact: {e}")
# Fallback: store in state
tool_context.state[StateKeys.LAST_FIX_REPORT] = fix_report
return {
"status": "success",
"message": "Fix report saved to state",
"size": len(report_json)
}
except Exception as e:
logger.error(f"Tool: Failed to save fix report: {e}", exc_info=True)
return {
"status": "error",
"message": str(e)
}
مرحله 7: خط لوله ثابت کامل را آزمایش کنید
Time to see the entire loop in action.
👉 Start the system:
adk web code_review_assistant
After running the adk web command, you should see output in your terminal indicating that the ADK Web Server has started, similar to this:
+-----------------------------------------------------------------------------+
| ADK Web Server started |
| |
| For local testing, access at http://localhost:8000. |
+-----------------------------------------------------------------------------+
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
👉 Test Prompt:
Please analyze the following:
def dfs_search_v1(graph, start, target):
"""Find if target is reachable from start."""
visited = set()
stack = start
while stack:
current = stack.pop()
if current == target:
return True
if current not in visited:
visited.add(current)
for neighbor in graph[current]:
if neighbor not in visited:
stack.append(neighbor)
return False
First, submit the buggy code to trigger the review pipeline . After it identifies the flaws, you will ask the agent to "Please fix the code" which triggers the powerful, iterative fix pipeline .
1. The Initial Review (Finding the Flaws)
This is the first half of the process. The four-agent review pipeline analyzes the code, checks its style, and runs a generated test suite. It correctly identifies a critical AttributeError and other issues, delivering a verdict: the code is BROKEN , with a test pass rate of only 84.21% .
2. The Automated Fix (The Loop in Action)
This is the most impressive part. When you ask the agent to fix the code, it doesn't just make one change. It kicks off an iterative Fix and Validate Loop that works just like a diligent developer: it tries a fix, tests it thoroughly, and if it's not perfect, it tries again.
Iteration #1: The First Attempt (Partial Success)
- The Fix: The
CodeFixeragent reads the initial report and makes the most obvious corrections. It changesstack = starttostack = [start]and usesgraph.get()to preventKeyErrorexceptions. - The Validation: The
TestRunnerimmediately re-runs the full test suite against this new code. - The Result: The pass rate improves significantly to 88.89% ! The critical bugs are gone. However, the tests are so comprehensive that they reveal two new, subtle bugs (regressions) related to handling
Noneas a graph or non-list neighbor values. The system marks the fix as PARTIAL .
Iteration #2: The Final Polish (100% Success)
- The Fix: Because the loop's exit condition (100% pass rate) was not met, it runs again. The
CodeFixernow has more information—the two new regression failures. It generates a final, more robust version of the code that explicitly handles those edge cases. - The Validation: The
TestRunnerexecutes the test suite one last time against the final version of the code. - The Result: A perfect 100% pass rate . All original bugs and all regressions are resolved. The system marks the fix as SUCCESSFUL and the loop exits.
3. The Final Report: A Perfect Score
With a fully validated fix, the FixSynthesizer agent takes over to present the final report, transforming the technical data into a clear, educational summary.
متریک | قبل از | بعد از | بهبود |
Test Pass Rate | 84.21% | 100% | ▲ 15.79% |
Style Score | 88 / 100 | 98 / 100 | ▲ 10 pts |
Bugs Fixed | 0 از 3 | 3 از 3 | ✅ |
✅ The Final, Validated Code
Here is the complete, corrected code that now passes all 19 tests, demonstrating the successful fix:
def dfs_search_v1(graph, start, target):
"""Find if target is reachable from start."""
# Handles 'None' graph input
if graph is None:
return False
visited = set()
# Fixes the critical AttributeError
stack = [start]
while stack:
current = stack.pop()
if current == target:
return True
if current not in visited:
visited.add(current)
# Safely gets neighbors to prevent KeyError
neighbors = graph.get(current)
if neighbors is None:
continue
# Validates that neighbors are iterable
if not isinstance(neighbors, (list, set, tuple)):
raise TypeError(
f"Graph value for node '{current}' is of type "
f"{type(neighbors).__name__}. Expected a list, set, or tuple."
)
for neighbor in neighbors:
if neighbor not in visited:
stack.append(neighbor)
return False
👉💻 Once you're done testing, return to your Cloud Shell Editor terminal and press Ctrl+C to stop the ADK Dev UI.
What You've Built
You now have a complete automated fix pipeline that:
✅ Generates fixes - Based on review analysis
✅ Validates iteratively - Tests after each fix attempt
✅ Retries automatically - Up to 3 attempts for success
✅ Exits intelligently - Via escalate when successful
✅ Tracks improvements - Compares before/after metrics
✅ Provides artifacts - Downloadable fix reports
Key Concepts Mastered
LoopAgent vs Sequential:
- Sequential: One pass through agents
- LoopAgent: Repeats until exit condition or max iterations
- Exit via
tool_context.actions.escalate = True
State Evolution Across Iterations:
-
CODE_FIXESupdated each iteration - Test results show improvement over time
- Validator sees cumulative changes
Multi-Pipeline Architecture:
- Review pipeline: Read-only analysis (Module 5)
- Fix loop: Iterative correction (Module 6 inner loop)
- Fix pipeline: Loop + synthesizer (Module 6 outer)
- Root agent: Orchestrates based on user intent
Tools Controlling Flow:
-
exit_fix_loop()sets escalate - Any tool can signal loop completion
- Decouples exit logic from agent instructions
Max Iterations Safety:
- Prevents infinite loops
- Ensures system always responds
- Presents best attempt even if not perfect
بعد چه است
In the final module, you'll learn how to deploy your agent to production :
- Setting up persistent storage with VertexAiSessionService
- Deploying to Agent Engine on Google Cloud
- Monitoring and debugging production agents
- Best practices for scaling and reliability
You've built a complete multi-agent system with sequential and loop architectures. The patterns you've learned - state management, dynamic instructions, tool orchestration, and iterative refinement - are production-ready techniques used in real agentic systems.
7. Deploying to Production
مقدمه
Your code review assistant is now complete with review and fix pipelines working locally. The missing piece: it only runs on your machine. In this module, you'll deploy your agent to Google Cloud, making it accessible to your team with persistent sessions and production-grade infrastructure.
آنچه یاد خواهید گرفت:
- Three deployment paths: Local, Cloud Run, and Agent Engine
- Automated infrastructure provisioning
- Session persistence strategies
- Testing deployed agents
درک گزینه های استقرار
The ADK supports multiple deployment targets, each with different tradeoffs:
مسیرهای استقرار
عامل | Local ( | Cloud Run ( | Agent Engine ( |
پیچیدگی | حداقل | متوسط | کم |
تداوم جلسه | In-memory only (lost on restart) | Cloud SQL (PostgreSQL) | Vertex AI managed (automatic) |
زیرساخت | None (dev machine only) | Container + Database | Fully managed |
شروع سرد | N/A | 100-2000ms | 100-500ms |
مقیاس بندی | Single instance | Automatic (to zero) | خودکار |
Cost Model | Free (local compute) | Request-based + free tier | Compute-based |
UI Support | Yes (via | Yes (via | No (API only) |
بهترین برای | Development/testing | Variable traffic, cost control | Production agents |
Additional deployment option: Google Kubernetes Engine (GKE) is available for advanced users requiring Kubernetes-level control, custom networking, or multi-service orchestration. GKE deployment is not covered in this codelab but is documented in the ADK deployment guide .
What Gets Deployed
When deploying to Cloud Run or Agent Engine, the following is packaged and deployed:
- Your agent code (
agent.py, all sub-agents, tools) - Dependencies (
requirements.txt) - ADK API server (automatically included)
- Web UI (Cloud Run only, when
--with_uispecified)
Important differences:
- Cloud Run : Uses
adk deploy cloud_runCLI (builds container automatically) orgcloud run deploy(requires custom Dockerfile) - Agent Engine : Uses
adk deploy agent_engineCLI (no container building needed, directly packages Python code)
Step 1: Configure Your Environment
فایل .env خود را پیکربندی کنید
Your .env file (created in Module 3) needs updates for cloud deployment. Open .env and verify/update these settings:
Required for all cloud deployments:
# Your actual GCP Project ID (REQUIRED)
GOOGLE_CLOUD_PROJECT=your-project-id
# GCP region for deployments (REQUIRED)
GOOGLE_CLOUD_LOCATION=us-central1
# Use Vertex AI (REQUIRED)
GOOGLE_GENAI_USE_VERTEXAI=true
# Model configuration (already set)
WORKER_MODEL=gemini-2.5-flash
CRITIC_MODEL=gemini-2.5-pro
Set bucket names (REQUIRED before running deploy.sh):
The deployment script creates buckets based on these names. Set them now:
# Staging bucket for Agent Engine code uploads (REQUIRED for agent-engine)
STAGING_BUCKET=gs://your-project-id-staging
# Artifact storage for reports and fixed code (REQUIRED for both cloud-run and agent-engine)
ARTIFACT_BUCKET=gs://your-project-id-artifacts
Replace your-project-id with your actual project ID in both bucket names. The script will create these buckets if they don't exist.
Optional variables (created automatically if blank):
# Agent Engine ID (populated after first deployment)
AGENT_ENGINE_ID=
# Cloud Run Database credentials (created automatically if blank)
CLOUD_SQL_INSTANCE_NAME=
DB_USER=
DB_PASSWORD=
DB_NAME=
Authentication Check
If you encounter authentication errors during deployment:
gcloud auth application-default login
gcloud config set project $GOOGLE_CLOUD_PROJECT
Step 2: Understanding the Deployment Script
The deploy.sh script provides a unified interface for all deployment modes:
./deploy.sh {local|cloud-run|agent-engine}
Script Capabilities
Infrastructure provisioning:
- API enablement (AI Platform, Storage, Cloud Build, Cloud Trace, Cloud SQL)
- IAM permission configuration (service accounts, roles)
- Resource creation (buckets, databases, instances)
- Deployment with proper flags
- Post-deployment verification
Key Script Sections
- Configuration (lines 1-35) : Project, region, service names, defaults
- Helper Functions (lines 37-200) : API enablement, bucket creation, IAM setup
- Main Logic (lines 202-400) : Mode-specific deployment orchestration
Step 3: Prepare Agent for Agent Engine
Before deploying to Agent Engine, an agent_engine_app.py file is needed that wraps your agent for the managed runtime. This has been created for you already.
View code_review_assistant/agent_engine_app.py
👉 Open file:
"""
Agent Engine application wrapper.
This file prepares the agent for deployment to Vertex AI Agent Engine.
"""
from vertexai import agent_engines
from .agent import root_agent
# Wrap the agent in an AdkApp object for Agent Engine deployment
app = agent_engines.AdkApp(
agent=root_agent,
enable_tracing=True,
)
مرحله 4: استقرار به Agent Engine
Agent Engine is the recommended production deployment for ADK agents because it provides:
- Fully managed infrastructure (no containers to build)
- Built-in session persistence via
VertexAiSessionService - Automatic scaling from zero
- Cloud Trace integration enabled by default
How Agent Engine Differs from Other Deployments
زیر کاپوت،
deploy.sh agent-engine
موارد استفاده:
adk deploy agent_engine \
--project=$GOOGLE_CLOUD_PROJECT \
--region=$GOOGLE_CLOUD_LOCATION \
--staging_bucket=$STAGING_BUCKET \
--display_name="Code Review Assistant" \
--trace_to_cloud \
code_review_assistant
This command:
- Packages your Python code directly (no Docker build)
- Uploads to the staging bucket you specified in
.env - Creates a managed Agent Engine instance
- Enables Cloud Trace for observability
- Uses
agent_engine_app.pyto configure the runtime
Unlike Cloud Run which containerizes your code, Agent Engine runs your Python code directly in a managed runtime environment, similar to serverless functions.
Run the Deployment
From your project root:
./deploy.sh agent-engine
مراحل استقرار
Watch the script execute these phases:
Phase 1: API Enablement
✓ aiplatform.googleapis.com
✓ storage-api.googleapis.com
✓ cloudbuild.googleapis.com
✓ cloudtrace.googleapis.com
Phase 2: IAM Setup
✓ Getting project number
✓ Granting Storage Object Admin
✓ Granting AI Platform User
✓ Granting Cloud Trace Agent
Phase 3: Staging Bucket
✓ Creating gs://your-project-id-staging
✓ Setting permissions
Phase 4: Artifact Bucket
✓ Creating gs://your-project-id-artifacts
✓ Configuring access
Phase 5: Validation
✓ Checking agent.py exists
✓ Verifying root_agent defined
✓ Checking agent_engine_app.py exists
✓ Validating requirements.txt
Phase 6: Build & Deploy
✓ Packaging agent code
✓ Uploading to staging bucket
✓ Creating Agent Engine instance
✓ Configuring session persistence
✓ Setting up Cloud Trace integration
✓ Running health checks
This process takes 5-10 minutes as it packages the agent and deploys it to Vertex AI infrastructure.
Save Your Agent Engine ID
Upon successful deployment:
✅ Deployment successful!
Agent Engine ID: 7917477678498709504
Resource Name: projects/123456789/locations/us-central1/reasoningEngines/7917477678498709504
Endpoint: https://us-central1-aiplatform.googleapis.com/v1/...
⚠️ IMPORTANT: Save this in your .env file:
AGENT_ENGINE_ID=7917477678498709504
خود را به روز کنید
.env
file immediately:
echo "AGENT_ENGINE_ID=7917477678498709504" >> .env
This ID is required for:
- Testing the deployed agent
- Updating the deployment later
- Accessing logs and traces
What Was Deployed
Your Agent Engine deployment now includes:
✅ Complete review pipeline (4 agents)
✅ Complete fix pipeline (loop + synthesizer)
✅ All tools (AST analysis, style checking, artifact generation)
✅ Session persistence (automatic via VertexAiSessionService )
✅ State management (session/user/lifetime tiers)
✅ Observability (Cloud Trace enabled)
✅ Auto-scaling infrastructure
Step 5: Test Your Deployed Agent
Update Your .env File
After deployment, verify your .env includes:
AGENT_ENGINE_ID=7917477678498709504 # From deployment output
GOOGLE_CLOUD_PROJECT=your-project-id
GOOGLE_CLOUD_LOCATION=us-central1
اسکریپت تست را اجرا کنید
The project includes tests/test_agent_engine.py specifically for testing Agent Engine deployments:
python tests/test_agent_engine.py
آنچه آزمایش انجام می دهد
- Authenticates with your Google Cloud project
- Creates a session with the deployed agent
- Sends a code review request (the DFS bug example)
- Streams the response back via Server-Sent Events (SSE)
- Verifies session persistence and state management
Expected Output
Authenticated with project: your-project-id
Targeting Agent Engine: projects/.../reasoningEngines/7917477678498709504
Creating new session...
Created session: 4857885913439920384
Sending query to agent and streaming response:
data: {"content": {"parts": [{"text": "I'll analyze your code..."}]}}
data: {"content": {"parts": [{"text": "**Code Structure Analysis**\n..."}]}}
data: {"content": {"parts": [{"text": "**Style Check Results**\n..."}]}}
data: {"content": {"parts": [{"text": "**Test Results**\n..."}]}}
data: {"content": {"parts": [{"text": "**Final Feedback**\n..."}]}}
Stream finished.
چک لیست تایید
- ✅ Full review pipeline executes (all 4 agents)
- ✅ Streaming response shows progressive output
- ✅ Session state persists across requests
- ✅ No authentication or connection errors
- ✅ Tool calls execute successfully (AST analysis, style checking)
- ✅ Artifacts are saved (grading report accessible)
Alternative: Deploy to Cloud Run
While Agent Engine is recommended for streamlined production deployment, Cloud Run offers more control and supports the ADK web UI. This section provides an overview.
When to Use Cloud Run
Choose Cloud Run if you need:
- The ADK web UI for user interaction
- Full control over the container environment
- Custom database configurations
- Integration with existing Cloud Run services
How Cloud Run Deployment Works
زیر کاپوت،
deploy.sh cloud-run
موارد استفاده:
adk deploy cloud_run \
--project=$GOOGLE_CLOUD_PROJECT \
--region=$GOOGLE_CLOUD_LOCATION \
--service_name="code-review-assistant" \
--app_name="code_review_assistant" \
--port=8080 \
--with_ui \
--artifact_service_uri="gs://$ARTIFACT_BUCKET" \
--trace_to_cloud \
code_review_assistant
This command:
- Builds a Docker container with your agent code
- Pushes to Google Artifact Registry
- Deploys as a Cloud Run service
- Includes the ADK web UI (
--with_ui) - Configures Cloud SQL connection (added by script after initial deployment)
The key difference from Agent Engine: Cloud Run containerizes your code and requires a database for session persistence, while Agent Engine handles both automatically.
Cloud Run Deployment Command
./deploy.sh cloud-run
What's Different
زیرساخت:
- Containerized deployment (Docker built automatically by ADK)
- Cloud SQL (PostgreSQL) for session persistence
- Database auto-created by script or uses existing instance
Session Management:
- Uses
DatabaseSessionServiceinstead ofVertexAiSessionService - Requires database credentials in
.env(or auto-generated) - State persists in PostgreSQL database
UI Support:
- Web UI available via
--with_uiflag (handled by script) - Access at:
https://code-review-assistant-xyz.a.run.app
What You've Accomplished
Your production deployment includes:
✅ Automated provisioning via deploy.sh script
✅ Managed infrastructure (Agent Engine handles scaling, persistence, monitoring)
✅ Persistent state across all memory tiers (session/user/lifetime)
✅ Secure credential management (automatic generation and IAM setup)
✅ Scalable architecture (zero to thousands of concurrent users)
✅ Built-in observability (Cloud Trace integration enabled)
✅ Production-grade error handling and recovery
Key Concepts Mastered
Deployment Preparation:
-
agent_engine_app.py: Wraps agent withAdkAppfor Agent Engine -
AdkAppautomatically configuresVertexAiSessionServicefor persistence - Tracing enabled via
enable_tracing=True
Deployment Commands:
-
adk deploy agent_engine: Packages Python code, no containers -
adk deploy cloud_run: Builds Docker container automatically -
gcloud run deploy: Alternative with custom Dockerfile
Deployment Options:
- Agent Engine: Fully managed, fastest to production
- Cloud Run: More control, supports web UI
- GKE: Advanced Kubernetes control (see GKE deployment guide )
Managed Services:
- Agent Engine handles session persistence automatically
- Cloud Run requires database setup (or auto-created)
- Both support artifact storage via GCS
Session Management:
- Agent Engine:
VertexAiSessionService(automatic) - Cloud Run:
DatabaseSessionService(Cloud SQL) - Local:
InMemorySessionService(ephemeral)
Your Agent Is Live
Your code review assistant is now:
- Accessible via HTTPS API endpoints
- Persistent with state surviving restarts
- Scalable to handle team growth automatically
- Observable with complete request traces
- Maintainable through scripted deployments
بعد چه می شود؟ In Module 8, you'll learn to use Cloud Trace to understand your agent's performance, identify bottlenecks in the review and fix pipelines, and optimize execution times.
8. Production Observability
مقدمه
Your code review assistant is now deployed and running in production on Agent Engine. But how do you know it's working well? Can you answer these critical questions:
- Is the agent responding quickly enough?
- Which operations are slowest?
- Are the fix loops completing efficiently?
- Where are performance bottlenecks?
Without observability, you're operating blind. The --trace-to-cloud flag you used during deployment automatically enabled Cloud Trace, giving you complete visibility into every request your agent processes.
In this module, you'll learn to read traces, understand your agent's performance characteristics, and identify areas for optimization.
درک ردپاها و دهانه ها
What is a Trace?
A trace is the complete timeline of your agent handling a single request. It captures everything from when a user sends a query until the final response is delivered. Each trace shows:
- Total duration of the request
- All operations that executed
- How operations relate to each other (parent-child relationships)
- When each operation started and ended
اسپان چیست؟
A span represents a single unit of work within a trace. Common span types in your code review assistant:
-
agent_run: Execution of an agent (root agent or sub-agent) -
call_llm: Request to a language model -
execute_tool: Tool function execution -
state_read/state_write: State management operations -
code_executor: Running code with tests
Spans have:
- Name : What operation this represents
- Duration : How long it took
- Attributes : Metadata like model name, token counts, inputs/outputs
- Status : Success or failure
- Parent/child relationships : Which operations triggered which
ابزار دقیق
When you deployed with --trace-to-cloud , ADK automatically instruments:
- Every agent invocation and sub-agent call
- All LLM requests with token counts
- Tool executions with inputs/outputs
- State operations (read/write)
- Loop iterations in your fix pipeline
- Error conditions and retries
No code changes required - tracing is built into ADK's runtime.
Step 1: Access Cloud Trace Explorer
Open Cloud Trace in your Google Cloud Console:
- Navigate to Cloud Trace Explorer
- Select your project from the dropdown (should be pre-selected)
- You should see traces from your test in Module 7
If you don't see traces yet:
The test you ran in Module 7 should have generated traces. If the list is empty, generate some trace data:
python tests/test_agent_engine.py
Wait 1-2 minutes for traces to appear in the console.
What You're Looking At
The Trace Explorer shows:
- List of traces : Each row represents one complete request
- Timeline : When requests occurred
- Duration : How long each request took
- Request details : Timestamp, latency, span count
This is your production traffic log - every interaction with your agent creates a trace.
Step 2: Examine a Review Pipeline Trace
Click on any trace in the list to open the waterfall view
You'll see a Gantt chart showing the complete execution timeline. Here's what a typical review pipeline trace looks like:
invocation (2.3s) ────────────────────────────────────────────►
├── agent_run: CodeReviewAssistant (2.2s) ──────────────────►
│ ├── state_read: CODE_TO_REVIEW (0.01s) ►
│ ├── agent_run: CodeReviewPipeline (2.1s) ─────────────►
│ │ ├── agent_run: CodeAnalyzer (0.3s) ──────►
│ │ │ ├── execute_tool: analyze_code_structure (0.1s) ──►
│ │ │ └── call_llm: gemini-2.5-flash (0.15s) ────►
│ │ ├── agent_run: StyleChecker (0.2s) ──────►
│ │ │ ├── execute_tool: check_code_style (0.1s) ──►
│ │ │ └── call_llm: gemini-2.5-flash (0.08s) ──►
│ │ ├── agent_run: TestRunner (1.2s) ─────────────►
│ │ │ └── code_executor: BuiltInCodeExecutor (0.9s) ────►
│ │ └── agent_run: FeedbackSynthesizer (0.4s) ────────►
│ │ └── call_llm: gemini-2.5-flash (0.28s) ────►
Reading the Waterfall
Each bar represents a span. The horizontal position shows when it started, and the length shows how long it took.
Key insights from this trace:
- Total latency : 2.3 seconds from request to response
- Critical path : TestRunner takes 1.2s (52% of total time)
- Bottleneck : Code execution within TestRunner takes 0.9s (75% of TestRunner's time)
- State operations : Very fast (10ms each) - not a concern
- Pipeline structure : Sequential execution - CodeAnalyzer → StyleChecker → TestRunner → FeedbackSynthesizer
Inspecting Span Details
بر روی کلیک کنید
call_llm: gemini-2.5-flash
span under FeedbackSynthesizer
You'll see detailed attributes for this LLM call:
{
"name": "call_llm",
"span_kind": "LLM",
"duration": "280ms",
"attributes": {
"llm.model": "models/gemini-2.5-flash",
"llm.request_type": "GenerateContent",
"llm.usage.prompt_tokens": 845,
"llm.usage.completion_tokens": 234,
"llm.usage.total_tokens": 1079,
"llm.response.finish_reason": "STOP",
"status_code": "OK"
}
}
This shows:
- Which model was used
- How many tokens were consumed (input + output)
- Request duration
- Success/failure status
- The full prompt is also visible in the attributes (scroll to see it)
Understanding the Pipeline Flow
Notice how the trace reveals your architecture:
- Root agent (CodeReviewAssistant) receives the request
- State read retrieves the code to review
- Review pipeline orchestrates four sub-agents sequentially
- Each sub-agent uses tools and LLM calls to complete its work
- Final response flows back up through the hierarchy
This visibility helps you understand exactly what happens during each request.
Step 3: Analyze a Fix Pipeline Trace
The fix pipeline is more complex because it includes loops. Let's examine how traces capture iterative behavior.
Find a trace that includes "CodeFixPipeline" in the span names
You may need to scroll through your traces or send a request that triggers the fix pipeline. If you don't have one, you can generate it:
# In your test script, respond "yes" when asked to fix issues
python tests/test_agent_engine.py
Examining Loop Structure
Here's what a fix pipeline trace with 2 iterations looks like:
agent_run: CodeFixPipeline (8.5s) ───────────────────────►
├── agent_run: FixAttemptLoop (7.8s) ───────────────────►
│ ├── loop_iteration: 1 (3.2s) ──────────►
│ │ ├── agent_run: CodeFixer (0.8s) ────►
│ │ │ └── call_llm: gemini-2.5-flash (0.7s) ───►
│ │ ├── agent_run: FixTestRunner (1.8s) ─────────►
│ │ │ └── code_executor: BuiltInCodeExecutor (1.5s) ─────►
│ │ └── agent_run: FixValidator (0.6s) ────►
│ │ ├── execute_tool: validate_fixed_style (0.2s) ──►
│ │ └── state_write: FIX_STATUS = "PARTIAL" ►
│ │
│ ├── loop_iteration: 2 (4.5s) ─────────────────►
│ │ ├── agent_run: CodeFixer (1.0s) ──────►
│ │ │ └── call_llm: gemini-2.5-flash (0.9s) ───►
│ │ ├── agent_run: FixTestRunner (2.0s) ────────►
│ │ │ └── code_executor: BuiltInCodeExecutor (1.7s) ─────►
│ │ └── agent_run: FixValidator (1.5s) ──────►
│ │ ├── execute_tool: compile_fix_report (0.3s) ──►
│ │ └── state_write: FIX_STATUS = "SUCCESSFUL" ►
│ │
│ └── loop_exit: escalation_triggered ►
│
└── agent_run: FixSynthesizer (0.7s) ────►
├── execute_tool: save_fix_report (0.2s) ──►
└── call_llm: gemini-2.5 (0.4s) ────►
مشاهدات کلیدی درباره حلقه ها
Iteration patterns:
- Two iterations : First attempt achieved partial success, second completed fully
- Progressive cost : Iteration 2 takes longer (4.5s vs 3.2s)
- State tracking : Each iteration writes FIX_STATUS to state
- Exit mechanism : Loop terminates via escalation when FIX_STATUS = "SUCCESSFUL"
What this reveals:
- Your loop architecture is working correctly
- Most fixes complete in 1-2 iterations (good design)
- Each iteration includes: fix generation → testing → validation
- Code execution dominates each iteration (1.5-1.7s)
- The loop properly exits when conditions are met
تفکیک هزینه:
- Iteration 1: 3.2s
- Iteration 2: 4.5s (longer due to accumulated context)
- Total loop: 7.8s
- Synthesis: 0.7s
- Total fix pipeline: 8.5s
Comparing to Review Pipeline
Review pipeline: ~2.3s
Fix pipeline: ~8.5s (with 2 iterations)
The fix pipeline takes ~3.7x longer, which makes sense:
- It includes iterative refinement
- It runs code multiple times (once per iteration)
- It accumulates context from previous attempts
Step 4: What You've Discovered
Performance Patterns
From examining traces, you now know:
Review pipeline:
- Typical duration: 2-3 seconds
- Main time consumer: TestRunner (code execution)
- LLM calls: Fast (100-300ms each)
- State operations: Negligible (10ms)
Fix pipeline:
- Typical duration: 4-5 seconds per iteration
- Most fixes: 1-2 iterations
- Code execution: 1.5-2.0 seconds per iteration
- Progressive cost: Later iterations take longer
What's fast:
- State reads/writes (10ms)
- Tool executions for analysis (100ms)
- Individual LLM calls (100-300ms)
What's slow (but necessary):
- Code execution with tests (0.9-2.0s)
- Multiple loop iterations (cumulative)
Where to Look for Issues
When reviewing traces in production, watch for:
- Unusually long traces (>15 seconds) - investigate what went wrong
- Failed spans (status != OK) - errors in execution
- Excessive loop iterations (>2) - fix quality issues
- Very high token counts - prompt optimization opportunities
What You've Learned
Through Cloud Trace, you now understand:
✅ Request flow : Complete execution path through your pipelines
✅ Performance characteristics : What's fast, what's slow, and why
✅ Loop behavior : How iterations execute and terminate
✅ Span hierarchy : How operations nest within each other
✅ Trace navigation : Reading waterfall charts effectively
✅ Token visibility : Where LLM costs accumulate
Key Concepts Mastered
Traces and Spans:
- Traces = complete request timelines
- Spans = individual operations within traces
- Waterfall view shows execution hierarchy
- Automatic instrumentation via ADK
Performance Analysis:
- Reading Gantt chart visualizations
- Identifying critical paths
- Understanding duration distributions
- Spotting bottlenecks
Production Visibility:
- Every operation traced automatically
- Token usage captured per LLM call
- State changes visible and traceable
- Loop iterations individually tracked
بعد چه است
Continue exploring Cloud Trace:
- Monitor traces regularly to catch issues early
- Compare traces to identify performance regressions
- Use trace data to inform optimization decisions
- Filter by duration to find slow requests
Advanced observability (optional):
- Export traces to BigQuery for complex analysis ( docs )
- Create custom dashboards in Cloud Monitoring
- Set up alerts for performance degradation
- Correlate traces with application logs
9. Conclusion: From Prototype to Production
What You've Built
You started with just seven lines of code and built a production-grade AI agent system:
# Where we started (7 lines)
agent = Agent(
model="gemini-2.5-flash",
instruction="Review Python code for issues"
)
# Where we ended (production system)
- Two distinct multi-agent pipelines (review and fix) built from 8 specialized agents.
- An iterative fix loop architecture for automated validation and retries.
- Real AST-based code analysis tools for deterministic, accurate feedback.
- Robust state management using the "constants pattern" for type-safe communication.
- Fully automated deployment to a managed, scalable cloud infrastructure.
- Complete, built-in observability with Cloud Trace for production monitoring.
Key Architectural Patterns Mastered
الگو | پیاده سازی | Production Impact |
یکپارچه سازی ابزار | AST analysis, style checking | Real validation, not just LLM opinions |
Sequential Pipelines | Review → Fix workflows | Predictable, debuggable execution |
Loop Architecture | Iterative fixing with exit conditions | Self-improving until success |
مدیریت دولتی | Constants pattern, three-tier memory | Type-safe, maintainable state handling |
استقرار تولید | Agent Engine via deploy.sh | Managed, scalable infrastructure |
قابلیت مشاهده | Cloud Trace integration | Full visibility into production behavior |
Production Insights from Traces
Your Cloud Trace data revealed critical insights:
✅ Bottleneck identified : TestRunner's LLM calls dominate latency
✅ Tool performance : AST analysis executes in 100ms (excellent)
✅ Success rate : Fix loops converge within 2-3 iterations
✅ Token usage : ~600 tokens per review, ~1800 for fixes
These insights drive continuous improvement.
Clean Up Resources (Optional)
If you're done experimenting and want to avoid charges:
Delete Agent Engine deployment:
import vertexai
client = vertexai.Client( # For service interactions via client.agent_engines
project="PROJECT_ID",
location="LOCATION",
)
RESOURCE_NAME = "projects/{PROJECT_ID}/locations/{LOCATION}/reasoningEngines/{RESOURCE_ID}"
client.agent_engines.delete(
name=RESOURCE_NAME,
force=True, # Optional, if the agent has resources (e.g. sessions, memory)
)
Delete Cloud Run service (if created):
gcloud run services delete code-review-assistant \
--region=$GOOGLE_CLOUD_LOCATION \
--quiet
Delete Cloud SQL instance (if created):
gcloud sql instances delete your-project-db \
--quiet
Clean up storage buckets:
gsutil -m rm -r gs://your-project-staging
gsutil -m rm -r gs://your-project-artifacts
مراحل بعدی
With your foundation complete, consider these enhancements:
- Add more languages : Extend tools to support JavaScript, Go, Java
- Integrate with GitHub : Automatic PR reviews
- Implement caching : Reduce latency for common patterns
- Add specialized agents : Security scanning, performance analysis
- Enable A/B testing : Compare different models and prompts
- Export metrics : Send traces to specialized observability platforms
خوراکی های کلیدی
- Start simple, iterate fast : Seven lines to production in manageable steps
- Tools over prompts : Real AST analysis beats "please check for bugs"
- State management matters : Constants pattern prevents typo bugs
- Loops need exit conditions : Always set max iterations and escalation
- Deploy with automation : deploy.sh handles all the complexity
- Observability is non-negotiable : You can't improve what you can't measure
منابع برای یادگیری مداوم
- ADK Documentation
- Advanced ADK Patterns
- Agent Engine Guide
- Cloud Run Documentation
- Cloud Trace Documentation
سفر شما ادامه دارد
You've built more than a code review assistant—you've mastered the patterns for building any production AI agent:
✅ Complex workflows with multiple specialized agents
✅ Real tool integration for genuine capabilities
✅ Production deployment with proper observability
✅ State management for maintainable systems
These patterns scale from simple assistants to complex autonomous systems. The foundation you've built here will serve you well as you tackle increasingly sophisticated agent architectures.
Welcome to production AI agent development. Your code review assistant is just the beginning.