CrewAI のマルチエージェントアプリケーションをデバッグする

agentpython() {
    logfile="/tmp/agentpython-stderr.log"
    python "$@" 2> >(tee "$logfile" >&2)
    if [[ -s "$logfile" ]]; then
        # If logfile is NOT empty, run check script
        python /Users/brettyoung/Desktop/dev25/tutorials/dbg_crw/debug_main.py "$logfile"
    else
        # If logfile is empty, clear it (truncate to zero length)
        > "$logfile"
    fi
}

profile_file=$(test -f ~/.zshrc && echo ~/.zshrc || (test -f ~/.bashrc && echo ~/.bashrc)); echo 'agentpython() {
    logfile="/tmp/agentpython-stderr.log"
    python "$@" 2> >(tee "$logfile" >&2)
    if [[ -s "$logfile" ]]; then
        python full_path_to_your_script "$logfile"
    else
        > "$logfile"
    fi
}' >> "$profile_file" && source "$profile_file" && echo "Added and sourced $profile_file"

import numpy as np

# Create a structured array
dt = np.dtype([('x', 'f8'), ('y', 'i4')])
arr = np.zeros(100, dtype=dt)

# Fill with data
arr['x'] = np.random.random(100)
arr['y'] = np.arange(100)

# Create problematic buffer view
buffer_data = arr.tobytes()[:-5]  # Truncated buffer

# This triggers a numpy buffer/memory bug
corrupted = np.frombuffer(buffer_data, dtype=np.complex128, count=-1)

# Try to use the corrupted array - this often segfaults
result = np.fft.fft(corrupted) * np.ones(len(corrupted))
print(f"Result shape: {result.shape}")

import os
import sys
import re
import requests
import tempfile
import webbrowser
import html
from pathlib import Path
from typing import Type, List, Optional, Dict, Any, Union
import json 

from pydantic import BaseModel, Field

from crewai import Agent, Task, Crew, Process
from crewai.tools import BaseTool
from crewai import BaseLLM, LLM
import weave; weave.init("crewai_debug_agent")

from langchain_openai import ChatOpenAI
import os
import re
from openai import OpenAI
from pydantic import BaseModel, Field
from typing import Type
import subprocess


LOGFILE = sys.argv[1] if len(sys.argv) > 1 else "/tmp/agentpython-stderr.log"
GITHUB_TOKEN = os.getenv('GITHUB_TOKEN')

# Read the log file BEFORE kicking off
LOG_CONTENT = ""
if os.path.exists(LOGFILE) and os.path.getsize(LOGFILE) > 0:
    with open(LOGFILE, 'r') as f:
        LOG_CONTENT = f.read()
    print(f"\033[95m[LOG CONTENT LOADED] {len(LOG_CONTENT)} characters from {LOGFILE}\033[0m")
else:
    LOG_CONTENT = "No log file found or file is empty"
    print(f"\033[95m[LOG] No content found in {LOGFILE}\033[0m")

def verbose_print(msg):
    print(f"\033[95m[LOG] {msg}\033[0m", flush=True)



# ----- Tool Input Schemas -----
class LogAnalysisInput(BaseModel):
    log_content: str = Field(..., description="Log content to analyze (already loaded)")

class SearchQueryInput(BaseModel):
    error_text: str = Field(..., description="Error text to generate search query from")

class CombinedSearchInput(BaseModel):
    query: str = Field(..., description="Search query for both GitHub issues and web")
    owner: str = Field(default="", description="GitHub repository owner")
    repo: str = Field(default="", description="GitHub repository name")

class FileAnalysisInput(BaseModel):
    log_content: str = Field(..., description="Log content to extract file information from")

class FileSnippetInput(BaseModel):
    file_path: str = Field(..., description="Path to the file to get snippet from")
    line: Optional[int] = Field(default=None, description="Line number to focus on")
    n_lines: int = Field(default=20, description="Number of lines to return")

class ToolSuggestionInput(BaseModel):
    error_message: str = Field(..., description="Error message to analyze")
    code_snippet: str = Field(..., description="Code snippet related to the error")

class ReportGenerationInput(BaseModel):
    log: str = Field(..., description="Error log content")
    file_snippet: str = Field(default="", description="Relevant code snippet")
    tools: str = Field(default="", description="Tool recommendations")
    gh_results: str = Field(default="", description="GitHub search results")
    web_results: str = Field(default="", description="Web search results")

# ----- Tools -----

class LogReaderTool(BaseTool):
    name: str = Field(default="Log Reader")
    description: str = Field(default="Provides access to the pre-loaded log content")
    args_schema: Type[BaseModel] = LogAnalysisInput
    
    def _run(self, log_content: str = None) -> str:
        verbose_print(f"Using pre-loaded log content")
        
        if not LOG_CONTENT or LOG_CONTENT == "No log file found or file is empty":
            return "[LOG] Log file empty or not found. No action needed."
        
        is_python_error = "Traceback" in LOG_CONTENT or "Exception" in LOG_CONTENT or "Error" in LOG_CONTENT
        error_type = "Python Error" if is_python_error else "General Error"
        return f"Error Type: {error_type}\n\nLog Content:\n{LOG_CONTENT}"

class SearchQueryGeneratorTool(BaseTool):
    name: str = Field(default="Search Query Generator")
    description: str = Field(default="Generates optimized search queries from error messages")
    args_schema: Type[BaseModel] = SearchQueryInput
    
    def _run(self, error_text: str) -> str:
        verbose_print("Generating search query via LLM...")
        try:
            prompt = (
                "Given this error or question, write a concise search query to help the person find a solution online. "
                "Output only the query (no explanation):\n\n" + error_text
            )
            query = llm.call(prompt)
            return f"Generated search query: {query.strip()}"
        except Exception as e:
            return f"Error generating search query: {str(e)}"




class CombinedSearchTool(BaseTool):
    name: str = Field(default="Combined GitHub & Web Search")
    description: str = Field(default="Searches both GitHub issues and the web in one call, returning both results.")
    args_schema: Type[BaseModel] = CombinedSearchInput

    def _run(self, query: str, owner: str = "", repo: str = "") -> dict:
        github_results = self._github_search(query, owner, repo)
        web_results = self._web_search(query)
        return {
            "github_issues": github_results,
            "web_search": web_results
        }

    def _github_search(self, query: str, owner: str, repo: str):
        import httpx
        GITHUB_TOKEN = os.getenv('GITHUB_TOKEN')
        url = 'https://api.github.com/search/issues'
        headers = {'Accept': 'application/vnd.github.v3+json'}
        if GITHUB_TOKEN:
            headers['Authorization'] = f'token {GITHUB_TOKEN}'
        gh_query = f'repo:{owner}/{repo} is:issue {query}' if owner and repo else query
        params = {'q': gh_query, 'per_page': 5}
        try:
            with httpx.Client(timeout=15) as client:
                resp = client.get(url, headers=headers, params=params)
                if resp.status_code == 200:
                    items = resp.json().get("items", [])
                    return [
                        {
                            "number": item.get("number"),
                            "title": item.get("title"),
                            "url": item.get("html_url"),
                            "body": (item.get("body") or "")[:500]
                        }
                        for item in items
                    ]
                else:
                    return [{"error": f"GitHub search failed: {resp.status_code} {resp.text}"}]
        except Exception as e:
            return [{"error": f"Error searching GitHub: {str(e)}"}]

    def _extract_json(self, text):
        m = re.search(r"```json\s*(.*?)\s*```", text, re.DOTALL)
        if not m:
            m = re.search(r"```(.*?)```", text, re.DOTALL)
        block = m.group(1) if m else text
        try:
            j = json.loads(block)
            return j if isinstance(j, list) else [j]
        except Exception:
            return []
        
    def _web_search(self, query: str, n_results: int = 5):
        # Your actual OpenAI-based tool call here
        from openai import OpenAI  # or however your actual OpenAI client is imported
        client = OpenAI()
        prompt = (
            f"Show me {n_results} of the most important/useful web results for this search along with a summary of the problem and proposed solution: '{query}'. "
            "Return as markdown JSON:\n"
            "[{\"title\": ..., \"url\": ..., \"date_published\": ..., \"snippet\": ...}]"
        )
        
        response = client.responses.create(
            model="gpt-4.1",  # or "gpt-4.1", or your available web-enabled model
            tools=[{"type": "web_search_preview"}],
            input=prompt,
        )
        return self._extract_json(response.output_text)




class FileAnalysisTool(BaseTool):
    name: str = Field(default="File Analysis")
    description: str = Field(default="Extracts file paths and line numbers from error logs")
    args_schema: Type[BaseModel] = FileAnalysisInput
    
    def _run(self, log_content: str = None) -> str:
        verbose_print("Invoking LLM to identify files from log...")
        # Use the global LOG_CONTENT if log_content not provided
        content_to_analyze = log_content or LOG_CONTENT
        try:
            prompt = (
                "Given this error message or traceback, list all file paths (and, if available, line numbers) involved in the error. "
                "Output one JSON per line, as:\n"
                '{"file": "path/to/file.py", "line": 123}\n'
                'If line is not found, use null.\n'
                f"\nError:\n{content_to_analyze}"
            )
            output = llm.call(prompt)
            results = []
            for l in output.splitlines():
                l = l.strip()
                if not l:
                    continue
                try:
                    results.append(eval(l, {"null": None}))
                except Exception as exc:
                    verbose_print(f"[File Extraction Skipped Line]: {l!r} ({exc})")
            return f"Files found in error: {results}"
        except Exception as e:
            return f"Error analyzing files: {str(e)}"

class FileSnippetTool(BaseTool):
    name: str = Field(default="File Snippet Extractor")
    description: str = Field(default="Extracts code snippets from files around specific lines")
    args_schema: Type[BaseModel] = FileSnippetInput
    
    def _run(self, file_path: str, line: Optional[int] = None, n_lines: int = 20) -> str:
        if not os.path.exists(file_path):
            return f"File not found: {file_path}"
        try:
            with open(file_path, "r") as f:
                lines = f.readlines()
            if line and 1 <= line <= len(lines):
                s = max(0, line-6)
                e = min(len(lines), line+5)
                code = lines[s:e]
            else:
                code = lines[:n_lines]
            return f"Code snippet from {file_path}:\n{''.join(code)}"
        except Exception as e:
            return f"Error reading file {file_path}: {str(e)}"

class ToolSuggestionTool(BaseTool):
    name: str = Field(default="Tool Suggestion")
    description: str = Field(default="Suggests which debugging tools to use next based on error analysis")
    args_schema: Type[BaseModel] = ToolSuggestionInput
    
    def _run(self, error_message: str, code_snippet: str) -> str:
        verbose_print("Requesting tool suggestions via LLM...")
        prompt = (
            "You are an AI debugging orchestrator. The following is a Python error message and a snippet of code "
            "from a file involved in the error. Based on this, choose which tools should be used next, and explain why. "
            "Possible tools: github_issue_search, web_search. "
            "Always recommend github_issue_search as it's very helpful. "
            "Provide your recommendation in a clear, structured format.\n"
            "Error:\n" + error_message + "\n\nFile snippet:\n" + code_snippet
        )
        try:
            return llm.call(prompt).strip()
        except Exception as e:
            return f"Error generating tool suggestions: {str(e)}"

class ReportGeneratorTool(BaseTool):
    name: str = Field(default="HTML Report Generator")
    description: str = Field(default="Generates HTML debug reports")
    args_schema: Type[BaseModel] = ReportGenerationInput
    
    def _run(self, log: str, file_snippet: str = "", tools: str = "", gh_results: str = "", web_results: str = "") -> str:
        verbose_print("Writing HTML report ...")
        out_path = os.path.join(tempfile.gettempdir(), 'dbg_report.html')
        try:
            with open(out_path, "w", encoding="utf-8") as f:
                f.write("<html><head><meta charset='utf-8'><title>Debug Results</title></head><body>\n")
                f.write("<h1 style='color:#444;'>Debugging Session Report</h1>\n")
                f.write("<h2>Error Log</h2>")
                f.write("<pre style='background:#f3f3f3;padding:8px;'>" + html.escape(log or "None") + "</pre>")
                if file_snippet:
                    f.write("<h2>Relevant Source Snippet</h2><pre style='background:#fafaff;padding:8px;'>" + html.escape(file_snippet) + "</pre>")
                if tools:
                    f.write("<h2>LLM Tool Recommendations</h2><pre style='background:#eef;'>" + html.escape(tools) + "</pre>")
                if gh_results:
                    f.write("<h2>GitHub & Web Search Results</h2><pre>" + html.escape(gh_results) + "</pre>")
                if web_results:
                    f.write("<h2>Web Search AI Answer</h2><pre>" + html.escape(web_results) + "</pre>")
                f.write("</body></html>")
            return f"HTML report generated and opened at: {out_path}"
        except Exception as e:
            return f"Error generating HTML report: {str(e)}"

# --- Tool Instances
log_reader_tool = LogReaderTool()
search_query_generator_tool = SearchQueryGeneratorTool()
combined_search_tool = CombinedSearchTool()
file_analysis_tool = FileAnalysisTool()
file_snippet_tool = FileSnippetTool()
tool_suggestion_tool = ToolSuggestionTool()
report_generator_tool = ReportGeneratorTool()


class CustomChatOpenAI(ChatOpenAI):
    def call(self, prompt, system_message=None):
        """
        Run inference on a prompt (string). Optionally provide a system message.
        
        Args:
            prompt (str): The user's message.
            system_message (str, optional): The system context for the assistant.
        
        Returns:
            str: The model's response content.
        """
        messages = []
        if system_message:
            messages.append(("system", system_message))
        messages.append(("human", prompt))
        result = self.invoke(messages)
        return result.content


llm = CustomChatOpenAI(model_name="gpt-4o-mini", temperature=0.3)
fouro_llm = CustomChatOpenAI(model_name="gpt-4o-mini", temperature=0.3)



# --- Agents ---
log_analyst_agent = Agent(
    role="Log Analysis Specialist",
    goal="Analyze the pre-loaded log content to identify errors and extract relevant information. Start by reading log with the log_reader_tool, then move on to usign the file_alaysis_tool to read important info from the file(s) involved in the error",
    backstory="Expert in parsing error logs and identifying the root causes of issues",
    tools=[log_reader_tool, file_analysis_tool],
    allow_delegation=False,
    llm=fouro_llm
)

search_specialist_agent = Agent(
    role="Search Query Specialist",
    goal="Generate optimized search queries from error messages for effective problem resolution. The search must be less that 100 chars long!!!!!!!!!!",
    backstory="Expert in crafting search queries that yield the most relevant debugging results. The search must be less that 100 chars long!!!!!!!!!!",
    tools=[search_query_generator_tool],
    allow_delegation=False,
    llm=fouro_llm
)


combined_research_agent = Agent(
    role="Combined Repository and Web Search Specialist",
    goal="Search both GitHub issues and the web for relevant solutions to errors and problems. You must use the combined_search_tool no matter what!!!!!!! Try to summarize each specific github/web problem and solution to help the user solve their issue. Make sure to include the links from the original sources next to their corresponding summaries / code etc",
    backstory="Expert in both GitHub open-source research and web documentation sleuthing for code solutions.",
    tools=[combined_search_tool],
    allow_delegation=False,
    llm=fouro_llm
)

code_analyst_agent = Agent(
    role="Code Analysis Specialist",
    goal="Analyze code snippets and suggest debugging approaches",
    backstory="Expert in code analysis and debugging strategy recommendation",
    tools=[file_snippet_tool],
    allow_delegation=False,
    llm=fouro_llm
)

report_generator_agent = Agent(
    role="Debug Report Generator",
    goal="Compile all debugging information into comprehensive HTML reports. Make sure to include the links to sources when they are provides -- but DO NOT make up links if they are not given.  Write an extensive report covering all possible solutions to the problem!!!",
    backstory="Specialist in creating detailed, actionable debugging reports",
    tools=[report_generator_tool],
    allow_delegation=False,
    llm=llm
)

# --- Tasks ---
log_analysis_task = Task(
    description=f"Analyze the pre-loaded log content. The log content is already available: {LOG_CONTENT[:500]}... Extract error information and identify the type of error.",
    expected_output="Detailed analysis of the log content including error type and content",
    agent=log_analyst_agent,
    output_file="log_analysis.md"
)

file_extraction_task = Task(
    description="Extract file paths and line numbers from the analyzed log content. Use the pre-loaded log content to identify which files are involved in the error.",
    expected_output="List of files and line numbers involved in the error",
    agent=log_analyst_agent,
    context=[log_analysis_task],
    output_file="file_analysis.md"
)

search_query_task = Task(
    description="Generate optimized search queries based on the error analysis for finding solutions online. The search must be less that 100 chars long!!!!!!!!!!",
    expected_output="Optimized search queries for the identified errors. The search must be less that 100 chars long!!!!!!!!!!",
    agent=search_specialist_agent,
    context=[log_analysis_task],
    output_file="search_queries.md"
)

combined_search_task = Task(
    description="Use the search queries to search both GitHub issues and the wide web for solutions. Make sure to make a very robust report incorporating ALL sources. Dont just give desciptions of the issue- write a detailed summary showcasing code and exact explanations to issues in the report.",
    expected_output="Relevant GitHub issues and web documentation/articles/answers.",
    agent=combined_research_agent,
    context=[search_query_task],
    output_file="combined_results.md"
)

code_analysis_task = Task(
    description="Extract and analyze code snippets from the implicated files. Suggest debugging tools and approaches.",
    expected_output="Code snippets and debugging tool recommendations",
    agent=code_analyst_agent,
    context=[file_extraction_task],
    output_file="code_analysis.md"
)

report_generation_task = Task(
    description="Compile all debugging information into a comprehensive HTML report and open it in the browser. Make sure to make a very robust report incorporating ALL sources Make sure to include the links to sources when they are provides -- but DO NOT make up links if they are not given. -- ALL sourced information must be cited!!!!!! Write an extensive report covering all possible solutions to the problem!!!",
    expected_output="Complete HTML debugging report",
    agent=report_generator_agent,
    context=[log_analysis_task, combined_search_task, code_analysis_task],
    output_file="debug_report.html"
)

# --- Run Crew ---
crew = Crew(
    agents=[
        log_analyst_agent,
        search_specialist_agent,
        combined_research_agent,
        code_analyst_agent,
        report_generator_agent
    ],
    tasks=[
        log_analysis_task,
        file_extraction_task,
        search_query_task,
        combined_search_task,
        code_analysis_task,
        report_generation_task
    ],
    process=Process.sequential,
    verbose=True
)

if __name__ == "__main__":
    print(f"\033[95m[STARTING] Log content loaded: {len(LOG_CONTENT)} chars\033[0m")
    result = crew.kickoff()
    print("\n\nDebug Analysis Complete:\n")
    print(result)
    
    # Try to open the generated report
    report_path = './debug_report.html'
    if os.path.exists(report_path):
        verbose_print(f"Opening final report: {report_path}")
        if sys.platform.startswith("darwin"):
            subprocess.Popen(['open', report_path])
        elif sys.platform.startswith("linux"):
            subprocess.Popen(['xdg-open', report_path])
        elif sys.platform.startswith("win"):
            os.startfile(report_path)


class CustomChatOpenAI(ChatOpenAI):
    def call(self, prompt, system_message=None):
        """
        Run inference on a prompt (string). Optionally provide a system message.
        
        Args:
            prompt (str): The user's message.
            system_message (str, optional): The system context for the assistant.
        
        Returns:
            str: The model's response content.
        """
        messages = []
        if system_message:
            messages.append(("system", system_message))
        messages.append(("human", prompt))
        result = self.invoke(messages)
        return result.content

from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse
import uvicorn
import os
from typing import List, Dict, Any, Optional, Union
import requests

# Your re-used settings
API_KEY = os.getenv("OPENROUTER_API_KEY") or "your_api_key"
SITE_URL = "https://your-site-url.com"
SITE_NAME = "Your Site Name"
MODEL = "qwen/qwen3-32b"

# Your OpenRouterCerebrasLLM class (truncated for brevity; copy your full code here)
class OpenRouterCerebrasLLM:
    def __init__(
        self,
        model: str,
        api_key: str,
        site_url: str,
        site_name: str,
        temperature: Optional[float] = None,
    ):
        self.model = model
        self.temperature = temperature
        self.api_key = api_key
        self.site_url = site_url
        self.site_name = site_name
        self.endpoint = "https://openrouter.ai/api/v1/chat/completions"
        
    def call(
        self,
        messages: Union[str, List[Dict[str, str]]],
        tools: Optional[List[dict]] = None,
        callbacks: Optional[List[Any]] = None,
        available_functions: Optional[Dict[str, Any]] = None,
    ) -> Union[str, Any]:
        if isinstance(messages, str):
            messages = [{"role": "user", "content": messages}]
        payload = {
            "model": self.model,
            "messages": messages,
            "temperature": self.temperature,
            "provider": {
                "order": ["cerebras"],
                "allow_fallbacks": False
            }
        }
        headers = {
            "Authorization": f"Bearer {self.api_key}",
            "HTTP-Referer": self.site_url,
            "X-Title": self.site_name,
            "Content-Type": "application/json"
        }
        response = requests.post(
            self.endpoint,
            headers=headers,
            json=payload,
            timeout=30
        )
        response.raise_for_status()
        result = response.json()
        return result

# Initialize the FastAPI app and the LLM once
app = FastAPI()
llm = OpenRouterCerebrasLLM(
    model=MODEL,
    api_key=API_KEY,
    site_url=SITE_URL,
    site_name=SITE_NAME,
    temperature=0.7
)

@app.post("/v1/chat/completions")
async def chat_completions(request: Request):
    body = await request.json()
    messages = body.get("messages")
    # you can also handle tools, temperature, etc. here
    try:
        raw_response = llm.call(messages)
        # This returns the full OpenRouter response. If you want to narrow down to only the OpenAI-compatible fields,
        # you could filter here, but for maximum compatibility just return as-is.
        return JSONResponse(raw_response)
    except Exception as e:
        return JSONResponse({"error": str(e)}, status_code=500)

if __name__ == "__main__":
    print(f"Serving local OpenAI-compatible LLM proxy on http://localhost:8001/v1/chat/completions")
    print(f"Forwarding all requests to: {MODEL}, via OpenRouter w/ your secret/settings")
    uvicorn.run(app, host="0.0.0.0", port=8001)