5.4 Codebase Exploration and Context Degradation

# Start broad, then zoom in
exploration_steps = [
    "ls -la",                           # Root structure
    "cat README.md",                    # Project overview
    "find . -name '*.py' | head -20",   # Main files
    "cat src/main.py",                  # Entry point
    # Then: follow imports, find relevant modules
]

# Start with what you're looking for
exploration_steps = [
    "grep -r 'def authenticate' .",     # Find the function
    "grep -r 'import auth' .",          # Find who uses it
    "cat src/auth/middleware.py",       # Read the implementation
]

# Understand architecture first
exploration_steps = [
    "find . -name '*.py' -path '*/models/*'",   # Data layer
    "find . -name '*.py' -path '*/routes/*'",   # API layer
    "find . -name '*.py' -path '*/services/*'", # Business logic
    "cat pyproject.toml",                        # Dependencies
]

# แทนที่จะ cat ทั้งไฟล์ (อาจ 1000+ lines)
# อ่านแค่ส่วนที่เกี่ยวข้อง
tools = [{
    "name": "read_file",
    "input_schema": {
        "properties": {
            "path": {"type": "string"},
            "start_line": {"type": "integer"},
            "end_line": {"type": "integer"}
        }
    }
}]

# Agent workflow
findings = []

# Read file 1
content = read_file("src/auth.py")
findings.append("auth.py: JWT-based auth, uses Redis for session store")

# Read file 2
content = read_file("src/database.py")  
findings.append("database.py: SQLAlchemy ORM, PostgreSQL, connection pooling")

# When context is getting full, work from summaries not raw files

# Level 1: File list
files = list_directory("src/")  # 50 files

# Level 2: Read only relevant files
relevant = [f for f in files if "auth" in f or "user" in f]

# Level 3: Read specific functions
for f in relevant:
    outline = get_function_signatures(f)  # Just signatures, not body
    if "login" in outline:
        full_content = read_file(f)  # Only read full when needed

# ใส่ข้อมูลสำคัญที่สุดเป็น message สุดท้าย
messages = [
    *old_context,                    # Less attention
    {"role": "user", "content": f"""
    Based on my exploration, the key findings are:
    {important_summary}              # More attention (recent)
    
    Now, please {actual_task}
    """}
]

# Write findings to a file — can re-read later
def save_exploration_notes(findings):
    with open(".agent-notes.md", "w") as f:
        f.write("# Codebase Exploration Notes\n\n")
        for finding in findings:
            f.write(f"- {finding}\n")

# Later: re-read only what's needed
def load_relevant_notes(topic):
    notes = read_file(".agent-notes.md")
    return filter_by_topic(notes, topic)

# Main agent doesn't read everything
# Delegates focused questions to sub-agents
answer = await agent(f"""
    In the codebase at {repo_path}:
    Find where user authentication happens.
    Return ONLY: file path, function name, and auth method used.
    Do not return full file contents.
""")