docs: merge docstrings and inline comments from agent

backend/agent/coding_agent.py

@@ -49,8 +49,15 @@ MAX_HISTORY_CHARS = 80_000
 # ═════════════════════════════════════════════════════════════════════════════
 
 def build_all_tool_description() -> str:
-    """Get relevant tools from the MCP servers based on the query."""
+    """Build a formatted string listing every registered MCP tool.
 
+    The returned string is embedded verbatim in the SYSTEM_PROMPT so the LLM
+    knows which tools exist and what arguments they expect.
+
+    Returns:
+        Newline-separated list of tool descriptions in the format
+        ``"- <tool_name>: <description>"``.
+    """
     all_tools = adapter.get_all_tools()
     print(f"Building tool description for {len(all_tools)} tools.")
 
@@ -58,17 +65,30 @@ def build_all_tool_description() -> str:
     for tool in all_tools:
         pprint.pprint(f"{tool}")
         descriptions.append(f"- {tool['tool_name']}: {tool['tool_description']}")
- 
+
     return "\n".join(descriptions)
 
 async def dispatch_tool(tool_name: str, arguments: dict) -> str:
-    """Call a tool by name with the given arguments using the MCP adapter."""
+    """Execute a named tool and return its output as a plain string.
 
+    Handles the special "done" pseudo-tool locally (it signals completion and
+    is never forwarded to an MCP server).  All other tools are forwarded to the
+    MCPToolAdapter which routes them to the correct MCP server process.
+
+    Args:
+        tool_name: Name of the tool to execute (e.g. "write_file", "done").
+        arguments: Dict of arguments for the tool.
+
+    Returns:
+        The tool's text output, a "DONE: ..." completion message, or an error
+        string beginning with "Tool error:" / "Error calling tool:" on failure.
+    """
     if tool_name == "done":
-        # Handle the "done" tool locally since it's not an MCP tool
+        # The "done" tool is a sentinel — it lives only in the agent protocol,
+        # not in any MCP server, so we resolve it directly here.
         summary = arguments.get("summary", "Task completed.")
         return f"DONE: {summary}"
-    
+
     try:
         print(f"Trying to call tool '{tool_name}' in dispatch_tool through MCPToolAdapter...")
         result = await adapter.call_tool(tool_name, arguments)
@@ -76,12 +96,14 @@ async def dispatch_tool(tool_name: str, arguments: dict) -> str:
         print(f"Raw result from tool '{tool_name}': {result}")
 
         if result.isError:
+            # MCP servers signal tool-level errors via the isError flag rather
+            # than raising exceptions, so we surface them explicitly.
             texts = [block.text for block in result.content if block.type == "text"]
             return f"Tool error: {' '.join(texts)}"
-        
+
         texts = [block.text for block in result.content if block.type == "text"]
-        return "\n".join(texts)   
-    
+        return "\n".join(texts)
+
     except Exception as e:
         return f"Error calling tool '{tool_name}': {e}"
 
@@ -167,23 +189,33 @@ def truncate_result(result: str) -> str:
 
 
 def trim_messages(messages: list) -> list:
-    """Drop old messages when history exceeds MAX_HISTORY_CHARS.
-    Always keeps the system prompt (index 0) and original task (index 1).
+    """Kürzt die Konversations-History wenn sie das Kontextfenster überschreitet.
+
+    Behält immer den System-Prompt (Index 0) und die ursprüngliche Aufgabe (Index 1).
+    Entfernt die ältesten Nachrichten zuerst und injiziert danach einen Erinnerungs-
+    Hinweis damit der Agent den Überblick behält.
+
+    Args:
+        messages: Vollständige Konversations-History als Liste von {role, content} Dicts.
+
+    Returns:
+        Gekürzte History mit maximal MAX_HISTORY_CHARS Zeichen, immer mit Head + Reminder + Tail.
     """
     total = sum(len(m["content"]) for m in messages)
     if total <= MAX_HISTORY_CHARS:
         return messages
 
+    # Protect the two anchor messages that must never be discarded.
     head = messages[:2]
     tail = messages[2:]
     original_task = messages[1]["content"] if len(messages) > 1 else ""
 
-    # Drop the oldest messages first (index 2 onwards) until we are under the limit.
-    # The system prompt (0) and original task (1) are never dropped.
+    # Drop the oldest non-anchor messages first until we are under the limit.
     while tail and sum(len(m["content"]) for m in head + tail) > MAX_HISTORY_CHARS:
         tail.pop(0)
 
-    # Inject a reminder so the agent doesn't lose track of its goal after trimming.
+    # After trimming, inject a reminder so the agent doesn't lose track of its goal.
+    # Without this the agent might restart the task or repeat work it already did.
     reminder = {
         "role": "user",
         "content": (
@@ -196,13 +228,22 @@ def trim_messages(messages: list) -> list:
     return head + [reminder] + tail
 
 def _repair_json_strings(text: str) -> str:
-    """
-    Replace unescaped control characters (newline, tab, carriage return)
-    inside JSON string values with their proper escape sequences.
+    """Replace unescaped control characters inside JSON string values.
 
-    LLMs frequently emit literal newlines inside long string values, which
-    is invalid JSON. This function fixes that without touching structural
-    whitespace outside strings.
+    LLMs frequently emit literal newlines, tabs, or carriage-returns inside
+    long string values (e.g. code content), which is invalid JSON.  This
+    function replaces those characters with their proper ``\\n`` / ``\\t`` /
+    ``\\r`` escape sequences without touching structural whitespace that lives
+    outside of string literals.
+
+    The parser is a simple state-machine that tracks whether the current
+    character is inside a quoted string, respecting backslash escapes.
+
+    Args:
+        text: Raw JSON text that may contain unescaped control characters.
+
+    Returns:
+        Repaired JSON text with control characters properly escaped inside strings.
     """
     result: list[str] = []
     in_string = False
@@ -210,6 +251,8 @@ def _repair_json_strings(text: str) -> str:
     _escapes = {'\n': '\\n', '\r': '\\r', '\t': '\\t'}
     for ch in text:
         if escape:
+            # The previous character was a backslash — emit this char literally
+            # and reset the escape flag.
             result.append(ch)
             escape = False
             continue
@@ -218,10 +261,12 @@ def _repair_json_strings(text: str) -> str:
             escape = True
             continue
         if ch == '"':
+            # Toggle string-mode on every unescaped double quote.
             in_string = not in_string
             result.append(ch)
             continue
         if in_string and ch in _escapes:
+            # Replace the bare control character with its escape sequence.
             result.append(_escapes[ch])
             continue
         result.append(ch)
@@ -285,13 +330,25 @@ def extract_json(text: str) -> str:
 
 
 def _strip_code_fences(text: str) -> str:
-    """Remove markdown code fences (```json ... ```) from a string."""
+    """Remove a single wrapping markdown code fence from a string.
+
+    Handles both `` ```json `` and plain `` ``` `` opening fences.  If the text
+    does not start with a fence the string is returned unchanged.
+
+    Args:
+        text: Raw LLM response that may be wrapped in a markdown code block.
+
+    Returns:
+        The text with the opening fence line and optional closing `` ``` `` line
+        removed, stripped of surrounding whitespace.
+    """
     if text is None:
         return ""
-    
+
     text = text.strip()
     if text.startswith("```"):
         lines = text.split("\n")
+        # Omit the last line only if it is a closing fence; otherwise keep everything.
         end = -1 if lines[-1].strip() == "```" else len(lines)
         text = "\n".join(lines[1:end])
     return text.strip()
@@ -328,11 +385,18 @@ class CodingAgent:
         self.api_key = os.getenv("API_KEY")
         self.model   = os.getenv("MODEL")
 
-    #async def _call_api(self, messages: list) -> str:
     def _call_api(self, messages: list) -> str:
+        """Send the conversation history to the LLM and return its reply text.
 
-        """Make a raw API call and return the response content string."""
-        
+        Args:
+            messages: Full conversation history as a list of {role, content} dicts.
+
+        Returns:
+            The raw string content of the assistant's reply.
+
+        Raises:
+            Exception: On non-200 HTTP status or an unexpected response schema.
+        """
         headers = {"Content-Type": "application/json"}
         if self.api_key and self.api_key != "EMPTY":
             headers["Authorization"] = f"Bearer {self.api_key}"
@@ -340,11 +404,13 @@ class CodingAgent:
         payload = {
             "model": self.model,
             "messages": messages,
-            "temperature": 0.2,   # low temperature → deterministic, more reliable tool calls
+            # Low temperature keeps the agent's tool selections deterministic and
+            # reduces the chance of hallucinated tool names or argument formats.
+            "temperature": 0.2,
             "max_tokens": 4096,
             "stream": False,
         }
-       
+
         response = requests.post(self.api_url, headers=headers, json=payload, timeout=60)
 
         if response.status_code != 200:

backend/agent/mcp_server_adapter.py

@@ -1,3 +1,15 @@
+"""Adapter layer between the CodingAgent and one or more MCP tool servers.
+
+MCPToolAdapter reads a JSON config file that lists MCP server processes, spawns
+each process via stdio, queries its available tools, and stores them in a flat
+registry.  At call time it re-spawns the appropriate server process, executes
+the requested tool, and returns the raw MCP result object.
+
+Design note: connections are opened per-call (not kept alive) because Streamlit
+reruns make it impractical to maintain long-lived async context managers across
+the synchronous/asynchronous boundary.
+"""
+
 import asyncio
 import json
 import sys
@@ -7,20 +19,42 @@ from pathlib import Path
 from mcp import ClientSession, StdioServerParameters
 from mcp.client.stdio import stdio_client
 
+
 class MCPToolAdapter:
+    """Discovers and dispatches MCP tools from one or more stdio-based MCP servers.
+
+    Workflow:
+        1. Call ``initialize_all_servers()`` once at startup to populate the
+           tool registry from every server listed in the config file.
+        2. Call ``get_all_tools()`` to retrieve the registry for building the
+           system-prompt tool description.
+        3. Call ``call_tool(name, arguments)`` whenever the agent wants to
+           execute a tool.  The adapter resolves the owning server, opens a
+           fresh connection, and returns the MCP result object.
+
+    Attributes:
+        config_path: Path (relative to this file) of the JSON server config.
+        servers: Dict mapping server name → raw config params dict.
+        tool_registry: Flat list of registered tool dicts, each containing
+            "server", "tool_name", and "tool_description".
+    """
+
     def __init__(self, config_path: str = "mcp_server_config.json"):
         self.config_path = config_path
         self.servers: Dict[str, Dict] = {}
-        #self.exit_stack: Dict[str, Any] = {}
         self.tool_registry: List[Dict[str, Any]] = []
 
     def _load_config(self) -> Dict[str, Any]:
-        """Lädt die Server-Konfiguration aus der JSON-Datei."""
+        """Load the MCP server configuration from the JSON file next to this module.
+
+        Returns:
+            Parsed config dict, or an empty dict if the file is missing or invalid.
+        """
         path = Path(__file__).parent / self.config_path
         if not path.exists():
             print(f"Config file not found: {path}")
             return {}
-        
+
         try:
             with open(path, 'r') as f:
                 return json.load(f)
@@ -29,7 +63,13 @@ class MCPToolAdapter:
             return {}
 
     async def initialize_all_servers(self):
-        """Lädt die Konfiguration und fragt alle Server ab, um die Tools zu registrieren."""
+        """Connect to every configured MCP server and register their tools.
+
+        Opens a short-lived stdio connection to each server, calls list_tools(),
+        and stores each discovered tool in ``self.tool_registry``.  Servers that
+        fail to connect are skipped with a warning so a single broken server does
+        not prevent the others from loading.
+        """
         print("Initializing MCP sessions...")
         config = self._load_config()
         print(f"Loaded config for servers: {list(config.keys())}")
@@ -39,18 +79,21 @@ class MCPToolAdapter:
 
             self.servers[server_name] = params
             server_script = str(Path(__file__).parent / params["args"][0])
+
+            # Always use the current Python interpreter so the server runs in the
+            # same virtual environment as the adapter, regardless of the literal
+            # command string in the config ("py", "python", "python3").
             if params.get("command") in ["py", "python", "python3"]:
                 server_command = sys.executable
             else:
                 server_command = params["command"]
-            
+
             server_params = StdioServerParameters(
                 command=server_command,
                 args=[server_script],
             )
-            
+
             try:
-               # Verbindung aufbauen
                 async with stdio_client(server_params) as (read_stream, write_stream):
                     print(f"Connected to {server_name}. Initializing session...")
                     async with ClientSession(read_stream, write_stream) as session:
@@ -60,9 +103,9 @@ class MCPToolAdapter:
                         print(f"Tools received from {server_name}: {result}")
                         tools = result.tools
                 print(f"Tools received from {server_name}: {result}")
-                #tools = getattr(result, 'tools', [])
 
                 for tool in tools:
+                    # Build a human-readable parameter description for the system prompt.
                     t_params = tool.inputSchema.get("properties", {})
                     if t_params:
                         param_lines = []
@@ -76,7 +119,6 @@ class MCPToolAdapter:
 
                     t_definition = f"- {tool.name}: {tool.description}\nParameters:\n{param_str}"
 
-
                     self.tool_registry.append({
                         "server": server_name,
                         "tool_name": tool.name,
@@ -84,22 +126,37 @@ class MCPToolAdapter:
                     })
 
                     print(f"Registered tool '{tool.name}' from {server_name}.")
-                
-                print(f"Session for {server_name} ready. {len(tools)} tools found.")
 
+                print(f"Session for {server_name} ready. {len(tools)} tools found.")
 
             except Exception as e:
                 print(f"Failed to initialize {server_name}: {e}")
 
     def get_all_tools(self) -> List[Dict[str, Any]]:
-        """Gibt alle gesammelten Tools zurück."""
+        """Return the full list of registered tools across all servers.
+
+        Returns:
+            List of dicts, each with keys "server", "tool_name", "tool_description".
+        """
         return self.tool_registry
 
     async def call_tool(self, tool_name: str, arguments: Dict[str, Any]):
-        """Findet den richtigen Server für ein Tool und führt es aus."""
-        # Suche in der Registry nach dem passenden Server
+        """Look up a tool in the registry, connect to its server, and execute it.
+
+        Opens a fresh stdio connection for every call.  This is intentionally
+        stateless so that server crashes or restarts are fully transparent.
+
+        Args:
+            tool_name: Name of the tool to call (must be in the registry).
+            arguments: Key-value arguments passed verbatim to the MCP server.
+
+        Returns:
+            The raw MCP ``CallToolResult`` object on success, or an error string
+            if the tool is not found or the server raises an exception.
+        """
+        # Look up which server owns this tool.
         tool_entry = next((t for t in self.tool_registry if t["tool_name"] == tool_name), None)
-        
+
         if not tool_entry:
             print(f"Tool '{tool_name}' not found in MCP adapter registry.")
             return f"Error: Tool '{tool_name}' not found in registry."
@@ -109,11 +166,12 @@ class MCPToolAdapter:
 
         if s_params:
             server_script = str(Path(__file__).parent / s_params["args"][0])
+            # Normalise the interpreter command the same way as in initialize_all_servers().
             if s_params.get("command") in ["py", "python", "python3"]:
                 server_command = sys.executable
             else:
                 server_command = s_params["command"]
-            
+
             server_params = StdioServerParameters(
                 command=server_command,
                 args=[server_script],
@@ -127,11 +185,16 @@ class MCPToolAdapter:
                 return result
             except Exception as e:
                 return f"Error calling tool '{tool_name}' on server '{server_name}': {str(e)}"
-        
+
         return f"Error: Session for server '{server_name}' not active."
 
     async def shutdown_all_sessions(self):
-        """Schließt alle offenen Verbindungen sauber."""
+        """Close all open server connections gracefully.
+
+        Note: This method references ``self.exit_stack`` which is not currently
+        populated (connections are opened per-call).  It is kept as a placeholder
+        for a future persistent-connection implementation.
+        """
         for server_name, (transport_gen, session) in self.exit_stack.items():
             try:
                 await session.__aexit__(None, None, None)

backend/agent/servers/mcp_server_code_execution.py

@@ -1,3 +1,19 @@
+"""MCP server that provides sandboxed Python code execution tools.
+
+Exposes the following MCP tools to the CodingAgent:
+  - analyse_structure   — AST-based structural summary of Python code
+  - lint_code           — pyflakes static analysis
+  - list_sandbox_packages — list packages installed in the sandbox venv
+  - install_package_into_sandbox — pip install into the sandbox venv
+  - reset_sandbox       — wipe and recreate the sandbox venv
+  - run_python_code_sandboxed — execute Python code inside the sandbox
+  - python_code_validation — syntax + safety check without execution
+
+The sandbox is an isolated virtual environment created on first use.
+All code submitted for execution is first checked by a static analyser that
+blocks dangerous imports and builtins before spawning any subprocess.
+"""
+
 import ast
 from datetime import datetime
 import subprocess
@@ -12,18 +28,30 @@ import shutil
 # ── Sandbox venv ────────────────────────────────────────────────────────────
 SERVER_BASE_DIR = Path(__file__).parent.resolve()
 SANDBOX_DIR = SERVER_BASE_DIR / ".mcp_sandbox"
+# Navigate four levels up from servers/ to reach the project root, then into workspace/.
 WORKSPACE_DIR = SERVER_BASE_DIR.parent.parent.parent.parent / "workspace"
 
+
 def get_sandbox_paths():
-    """Bestimmt die Executables innerhalb der Venv ohne os-Modul."""
+    """Locate (and lazily create) the sandbox venv and return its executable paths.
+
+    Creates the venv on first call if it does not yet exist.  Uses a simple
+    filesystem check to determine whether we are on Windows (``Scripts/``) or
+    POSIX (``bin/``), avoiding the ``os`` module which is blocked inside the
+    sandbox itself.
+
+    Returns:
+        Tuple of (python_exe_path, pip_exe_path) as strings.
+    """
     if not SANDBOX_DIR.exists():
         venv.create(SANDBOX_DIR, with_pip=True)
-    
-    bin_folder = "Scripts" if Path("C:/").exists() else "bin" # Einfacher Check für Windows
-    
+
+    # Path("C:/").exists() is True on Windows, False on Linux/macOS.
+    bin_folder = "Scripts" if Path("C:/").exists() else "bin"
+
     python_exe = SANDBOX_DIR / bin_folder / "python"
     pip_exe = SANDBOX_DIR / bin_folder / "pip"
-    
+
     return str(python_exe), str(pip_exe)
 
 PYTHON_EXE, PIP_EXE = get_sandbox_paths()
@@ -300,9 +328,17 @@ def install_package_into_sandbox(package_name: str) -> str:
 
 @mcp.tool()
 def reset_sandbox() -> str:
-    """Löscht die gesamte Sandbox und erstellt sie neu (Full Reset)."""
+    """Delete and recreate the sandbox virtual environment (full reset).
+
+    Useful when a package installation went wrong or the venv became corrupted.
+    Calling get_sandbox_paths() after deletion triggers the lazy creation logic.
+
+    Returns:
+        Confirmation string after the reset completes.
+    """
     if SANDBOX_DIR.exists():
         shutil.rmtree(SANDBOX_DIR)
+    # Re-calling get_sandbox_paths() triggers venv creation for the fresh sandbox.
     get_sandbox_paths()
     return "Sandbox wurde komplett zurückgesetzt."
 
@@ -325,22 +361,26 @@ def run_python_code_sandboxed(code: str) -> str:
         Combined stdout+stderr, or an error message in str format.
     """
 
+    # Reject code that references blocked modules or builtins before spawning a process.
     static_safety = check_code_safety(code)
     if static_safety:
         return f"Code rejected:{static_safety}"
 
+    # Each run gets its own temporary directory so concurrent runs don't interfere.
     run_id = datetime.now().strftime("%Y%m%d_%H%M%S")
     jail_dir = WORKSPACE_DIR / f"sandbox_run_{run_id}"
-    
+
     try:
         jail_dir.mkdir(parents=True, exist_ok=True)
 
+        # Minimal environment: only the sandbox Python is on PATH, HOME and TMPDIR
+        # point to the per-run jail directory so the subprocess cannot access user files.
         custom_env = {
-        "PYTHONPATH": str(WORKSPACE_DIR),
-        "PATH": str(Path(PYTHON_EXE).parent),
-        "HOME": str(jail_dir),
-        "TMPDIR": str(jail_dir)
-    }
+            "PYTHONPATH": str(WORKSPACE_DIR),
+            "PATH": str(Path(PYTHON_EXE).parent),
+            "HOME": str(jail_dir),
+            "TMPDIR": str(jail_dir)
+        }
 
         result = subprocess.run(
             [PYTHON_EXE, "-c", code],
@@ -349,47 +389,54 @@ def run_python_code_sandboxed(code: str) -> str:
             capture_output=True,
             text=True,
             timeout=EXEC_TIMEOUT)
-        
+
+        # Merge stdout and stderr so the agent sees all output in one block.
         output = result.stdout + result.stderr
 
         if len(output) > MAX_OUTPUT_LENGTH:
             output = output[:MAX_OUTPUT_LENGTH] + "\n...[output truncated]..."
-        
+
         if not output.strip():
             return "Code executed successfully (no output)."
-        
+
         return output
-    
+
     except subprocess.TimeoutExpired:
         return f"Error: Code execution exceeded time limit of {EXEC_TIMEOUT} seconds and was terminated."
     except Exception as e:
         return f"Error during code execution: {e}"
-    
+
     finally:
+        # Always clean up the per-run jail directory, even if execution failed.
         if jail_dir.exists():
             shutil.rmtree(jail_dir)
 
 
 @mcp.tool()
 def python_code_validation(code: str) -> str:
-    """
-    Validate Python code for syntax and safety without executing it.
-    This tool performs static analysis to check for syntax errors.
-    
+    """Validate Python code for syntax correctness and sandbox safety without executing it.
+
+    Performs two checks in sequence:
+      1. AST parsing to catch syntax errors.
+      2. check_code_safety() to detect blocked imports/builtins/path sequences.
+
     Args:
-        code: The Python code to validate in str format.
+        code: The Python source code to validate.
+
     Returns:
-        A message indicating the validation result.
-        And if sandboxed test execution is allowed.
+        A message indicating whether the code is valid and safe, or describing
+        the first violation found.  Returns None implicitly when the code is
+        both syntactically valid and safe (no safety concerns found).
     """
     try:
         ast.parse(code)
     except SyntaxError as e:
         return f"SyntaxError: {e}"
 
-    try: 
+    try:
         static_analysis_result = check_code_safety(code)
         if static_analysis_result:
+            # Inform the agent that the code would be rejected by run_python_code_sandboxed.
             return f"Valid Syntax, but with safety concerns: {static_analysis_result}; code execution is not allowed."
     except Exception as e:
         return f"Error during code safety analysis: {e}"

backend/agent/servers/mcp_server_file_search.py

@@ -1,10 +1,27 @@
+"""MCP server that provides file system read/write tools for the workspace directory.
+
+All operations are restricted to ALLOWED_DIR (the project workspace).  Paths
+that resolve outside this boundary are rejected with a ValueError so the agent
+cannot accidentally read or write arbitrary host-filesystem locations.
+
+Exposes the following MCP tools:
+  - list_files           — flat list of all workspace files
+  - get_file_tree        — tree-formatted directory listing
+  - search_files         — search file names and content
+  - read_file            — read a single file
+  - write_new_file       — create a new file (no overwrite)
+  - create_new_directory — create a new directory
+"""
+
 from pathlib import Path
 from mcp.server.fastmcp import FastMCP
 
 # ── Configuration ────────────────────────────────────────────────────────────
+# Navigate four levels up from servers/ to the project root, then into workspace/.
 project_dir = Path(__file__).resolve().parent.parent.parent.parent
 ALLOWED_DIR = project_dir / "workspace"
-ALLOWED_FILE_TYPES = [".py",".js",".html",".css",".json",".yaml",".yml",".sh",".md",".txt",".tex",".c",".cpp",".java"]
+ALLOWED_FILE_TYPES = [".py", ".js", ".html", ".css", ".json", ".yaml", ".yml",
+                      ".sh", ".md", ".txt", ".tex", ".c", ".cpp", ".java"]
 
 # ── Create the MCP server ────────────────────────────────────────────────────
 mcp = FastMCP("FileSearchServer")
@@ -63,12 +80,30 @@ def get_file_tree(dir_path: str=ALLOWED_DIR) -> str:
     
 
     def _tree(dir_path: Path, prefix="") -> str:
-        entries = sorted([e for e in dir_path.iterdir() if "__pycache__" not in e.parts], key=lambda x: (x.is_file(), x.name))
+        """Recursively build a tree string for the given directory.
+
+        Directories are sorted before files (``key=lambda x: (x.is_file(), x.name)``
+        puts dirs first because False < True).  __pycache__ entries are hidden.
+
+        Args:
+            dir_path: The directory to render.
+            prefix:   Indentation prefix accumulated during recursion.
+
+        Returns:
+            Multi-line string representing the subtree.
+        """
+        # Exclude __pycache__ at every level to keep output readable for the agent.
+        entries = sorted(
+            [e for e in dir_path.iterdir() if "__pycache__" not in e.parts],
+            key=lambda x: (x.is_file(), x.name)  # directories first, then files
+        )
         lines = []
         for i, entry in enumerate(entries):
+            # Use └── for the last entry to close the branch visually.
             connector = "└── " if i == len(entries) - 1 else "├── "
             lines.append(f"{prefix}{connector}{entry.name}")
             if entry.is_dir():
+                # Extend prefix with a blank column (last item) or │ (more items follow).
                 extension = "    " if i == len(entries) - 1 else "│   "
                 lines.append(_tree(entry, prefix + extension))
         return "\n".join(lines)

backend/agent/servers/mcp_server_web_search.py

@@ -1,3 +1,13 @@
+"""MCP server that provides web search and page-fetching tools.
+
+Exposes two MCP tools:
+  - web_search  — keyword search via DuckDuckGo, returns titles, URLs, snippets
+  - fetch_page  — fetch and extract readable text from a URL
+
+All outbound requests are guarded by _validate_url() which blocks non-HTTP
+schemes and private/loopback IP ranges to prevent SSRF vulnerabilities.
+"""
+
 from urllib.parse import urlparse
 from mcp.server.fastmcp import FastMCP
 
@@ -12,7 +22,22 @@ mcp = FastMCP("WebSearchServer")
 # ── Helper: URL validation (SSRF prevention) ─────────────────────────────────
 
 def _validate_url(url: str) -> str:
-    """Validate a URL to prevent SSRF attacks."""
+    """Validate a URL and raise ValueError if it could be used for an SSRF attack.
+
+    Blocks:
+    - Non-HTTP(S) schemes (file://, ftp://, etc.)
+    - Loopback and metadata addresses (localhost, 127.0.0.1, 169.254.169.254)
+    - RFC-1918 private IP ranges (10.x, 172.16-31.x, 192.168.x)
+
+    Args:
+        url: The URL string to validate.
+
+    Returns:
+        The original URL string unchanged if it passes all checks.
+
+    Raises:
+        ValueError: If the URL fails any of the security checks.
+    """
     parsed = urlparse(url)
 
     if parsed.scheme not in ("http", "https"):
@@ -22,10 +47,13 @@ def _validate_url(url: str) -> str:
 
     hostname = parsed.hostname or ""
 
+    # Block well-known loopback and cloud-metadata addresses.
     blocked_hosts = {"localhost", "127.0.0.1", "0.0.0.0", "169.254.169.254"}
     if hostname in blocked_hosts:
         raise ValueError(f"Blocked internal host: {hostname}")
 
+    # Block all private RFC-1918 ranges by checking the string prefix.
+    # This is a best-effort check; a full implementation would resolve DNS first.
     private_prefixes = (
         "10.", "172.16.", "172.17.", "172.18.", "172.19.",
         "172.20.", "172.21.", "172.22.", "172.23.", "172.24.",
@@ -100,6 +128,7 @@ def fetch_page(url: str) -> str:
 
         soup = BeautifulSoup(response.text, "html.parser")
 
+        # Remove boilerplate elements that add noise without informational value.
         for tag in soup(["script", "style", "nav", "footer"]):
             tag.decompose()
 

backend/managers/execution_engine.py

@@ -1,3 +1,10 @@
+"""Executes code files from the editor in isolated subprocesses.
+
+Supports Python (.py) via the system Python interpreter and LaTeX (.tex) via
+pdflatex.  All execution is time-bounded by RUN_TIMEOUT to prevent runaway
+processes from blocking the UI indefinitely.
+"""
+
 import subprocess
 from pathlib import Path
 

backend/managers/file_manager.py

@@ -12,6 +12,15 @@ WORKSPACE = Path("workspace")
 WORKSPACE.mkdir(exist_ok=True)
 
 class FileManager:
+    """Manages all file and folder operations inside the workspace directory.
+
+    Every public method resolves the given path and verifies that the result
+    stays within ``base_path`` before touching the filesystem.  This prevents
+    path-traversal attacks where a caller might pass ``../../etc/passwd``.
+
+    The workspace directory is created on first use if it does not yet exist.
+    """
+
     def __init__(self, base_path=Path("workspace")) -> None:
         self.base_path = Path(base_path)
         self.base_path.mkdir(exist_ok=True)

frontend/chat.py

@@ -10,10 +10,23 @@ from backend.managers.system_prompter import SystemPrompter
 
 # ── Agent Mode helpers ────────────────────────────────────────────────────────
 def _run_async(coro):
-    """Hilfsfunktion um async Code in sync Streamlit auszuführen"""
+    """Execute an async coroutine from synchronous Streamlit code.
+
+    Streamlit runs in a synchronous context, but the CodingAgent uses async
+    methods (for MCP tool calls).  This helper bridges the gap by reusing an
+    already-running event loop when one exists, or creating a new one otherwise.
+
+    Args:
+        coro: The coroutine to run.
+
+    Returns:
+        The return value of the coroutine.
+    """
     try:
+        # Reuse the loop that is already running (e.g. inside pytest-asyncio).
         loop = asyncio.get_running_loop()
     except RuntimeError:
+        # No running loop in this thread — create a fresh one.
         loop = asyncio.new_event_loop()
         asyncio.set_event_loop(loop)
     return loop.run_until_complete(coro)

frontend/editor.py

@@ -62,6 +62,14 @@ def _rename_dialog(file_path: str):
 
 @st.dialog("Delete File")
 def _delete_dialog(abs_file_path: str):
+    """Confirmation dialog before permanently deleting the given file.
+
+    Removes the file from disk and also cleans up the editor's open-files list,
+    content cache, and active_file pointer so the UI stays consistent.
+
+    Args:
+        abs_file_path: Absolute path to the file that should be deleted.
+    """
     fm = FileManager()
     file_name = Path(abs_file_path).name
     relative_path = str(Path(abs_file_path).relative_to(fm.base_path))

frontend/sidebar.py

@@ -53,6 +53,12 @@ def _delete_folder_dialog(folder_rel: str, folder_name: str):
 
 @st.dialog("Add File")
 def _add_file_dialog(parent_path: str = ""):
+    """Dialog for creating a new file inside the given folder (or workspace root).
+
+    Args:
+        parent_path: Workspace-relative path of the parent folder.  Pass an
+            empty string to create the file at the workspace root.
+    """
     with st.form("add_file_form"):
         name = st.text_input("File name:", placeholder="e.g. script.py")
         col1, col2 = st.columns(2)
@@ -77,6 +83,12 @@ def _add_file_dialog(parent_path: str = ""):
 
 @st.dialog("Add Folder")
 def _add_folder_dialog(parent_path: str = ""):
+    """Dialog for creating a new subfolder inside the given folder (or workspace root).
+
+    Args:
+        parent_path: Workspace-relative path of the parent folder.  Pass an
+            empty string to create the folder at the workspace root.
+    """
     with st.form("add_folder_form"):
         name = st.text_input("Folder name:", placeholder="e.g. utils")
         col1, col2 = st.columns(2)
@@ -166,6 +178,15 @@ def _rename_file_dialog(relative_file_path: str, file_name: str):
 
 @st.dialog("Delete File")
 def _delete_file_dialog(relative_file_path: str, file_name: str):
+    """Confirmation dialog before permanently deleting a file.
+
+    After a successful delete the file is also removed from the editor's
+    open-files list and content cache so it cannot be saved back to disk.
+
+    Args:
+        relative_file_path: Workspace-relative path to the file (used by FileManager).
+        file_name: Display name shown in the warning message.
+    """
     st.warning(f"Delete **{file_name}**? This cannot be undone.")
 
     col1, col2 = st.columns(2)
@@ -175,6 +196,7 @@ def _delete_file_dialog(relative_file_path: str, file_name: str):
                 abs_file_path = str(Path(fm.base_path) / relative_file_path)
                 st.session_state.open_files.remove(abs_file_path)
                 st.session_state.files_content.pop(abs_file_path, None)
+                # Fall back to the first remaining open file, or None if all tabs are closed.
                 if st.session_state.active_file == abs_file_path:
                     st.session_state.active_file = (
                         st.session_state.open_files[0]

frontend/state.py

@@ -10,45 +10,52 @@ from backend.managers.chat_manager import ChatManager
 
 
 def init_state():
-    # Sidebar state initialization
+    """Initialise all Streamlit session-state keys with safe defaults.
+
+    Uses ``if key not in st.session_state`` guards throughout so that existing
+    values are never overwritten on subsequent reruns — only missing keys are
+    set. This means it is safe to call multiple times per session.
+    """
+    # ── Sidebar state ─────────────────────────────────────────────────────────
+
     # last_selected tracks the previously clicked tree node to detect new clicks
+    # and avoid re-running the same file-open logic on every Streamlit rerender.
     if "last_selected" not in st.session_state:
         st.session_state.last_selected = None
 
-    # Absolute path and workspace-relative path of the currently highlighted folder
+    # Absolute path and workspace-relative path of the currently highlighted folder.
+    # Both are set together; both are cleared together when a folder is deselected.
     if "selected_folder" not in st.session_state:
         st.session_state.selected_folder = None
 
     if "selected_folder_rel" not in st.session_state:
         st.session_state.selected_folder_rel = None
 
-    # Chat manager (persists across reruns)
-    # ChatManager keeps the full conversation history in memory across reruns
+    # ── Chat manager ──────────────────────────────────────────────────────────
+
+    # ChatManager keeps the full conversation history in memory across reruns.
+    # Instantiated once and reused so history is not lost on page rerenders.
     if "chat_manager" not in st.session_state:
         st.session_state.chat_manager = ChatManager()
 
-    # Editor state initialization
-    # List of absolute file paths that are currently open as tabs
+    # ── Editor state ──────────────────────────────────────────────────────────
+
+    # Ordered list of absolute file paths currently open as editor tabs.
+    # The list order determines the visual tab order in the UI.
     if "open_files" not in st.session_state:
-        """A list of currently open file paths - absolute paths only. The order determines the tab order in the UI.
-        Format: [ "path/to/file1.py", "path/to/file2.js", ... ]
-        """
         st.session_state.open_files = []
 
-    # Dict mapping file path → current editor content (may be unsaved)
+    # Dict mapping absolute file path → current editor content (may differ from
+    # disk if the user has unsaved changes).
     if "files_content" not in st.session_state:
-        """A dictionary mapping file paths to their current content in the editor.
-        Format: { "path/to/file.py": "file content as string", ... }
-        """
         st.session_state.files_content = {}
 
-    # Absolute path of the file whose tab is currently active
+    # Absolute path of the file whose tab is currently active in the editor.
+    # Must always be one of the paths in open_files, or None if no file is open.
     if "active_file" not in st.session_state:
-        """The currently active file in the editor (absolute path in string e.g. "/workspace/path/to/file.py"). 
-        Should be one of the paths in open_files or None if no file is open."""
         st.session_state.active_file = None
 
-    # Index of the active tab (used by st.tabs)
+    # Index of the active tab — kept in sync with active_file for st.tabs().
     if "active_tab" not in st.session_state:
         st.session_state.active_tab = 0
 
@@ -58,33 +65,41 @@ def init_state():
     if "code_suggestions" not in st.session_state:
         st.session_state.code_suggestions = []
 
-    # Output dict from the last code run: {stdout, stderr, return_code}
+    # Output dict from the last code execution: {stdout, stderr, return_code}.
+    # Initialised as empty string so the editor view can safely check falsyness.
     if "code_execution_output" not in st.session_state:
         st.session_state.code_execution_output = ""
 
-    # Chat state initialization
-    # Flat list of {"role": ..., "content": ...} dicts shown as chat bubbles
+    # ── Chat state ────────────────────────────────────────────────────────────
+
+    # Flat list of {"role": ..., "content": ...} dicts rendered as chat bubbles.
+    # System messages are stored here too but skipped during display.
     if "chat_history" not in st.session_state:
         st.session_state.chat_history = []
 
-    # Agent Mode state
-    # Whether the UI is currently in Agent Mode (vs normal chat)
+    # ── Agent Mode state ──────────────────────────────────────────────────────
+
+    # Boolean toggle — True while the UI is in Coding Agent mode.
     if "agent_mode" not in st.session_state:
         st.session_state.agent_mode = False
 
-    # The live CodingAgent instance while a task is running
+    # The live CodingAgent instance while a task is running.
+    # Set by _start_agent(), cleared by _reset_agent().
     if "coding_agent" not in st.session_state:
         st.session_state.coding_agent = None
 
-    # Current status of the agent: "idle" | "waiting_approval" | "done"
+    # Lifecycle state of the agent: "idle" | "waiting_approval" | "done".
+    # Controls which sub-screen render_agent_mode() displays.
     if "agent_status" not in st.session_state:
         st.session_state.agent_status = "idle"
 
-    # List of completed steps shown in the collapsible Agent Log
+    # Chronological list of completed step records shown in the Agent Log expander.
+    # Each entry: {"thought": str, "tool": str, "arguments": dict, "result": str}
     if "agent_log" not in st.session_state:
         st.session_state.agent_log = []
 
-    # The action the agent proposed but has not yet been approved or rejected
+    # The action the agent has proposed but that has not yet been approved or
+    # rejected by the user.  Stored as the raw dict returned by propose_next_action().
     if "agent_pending_action" not in st.session_state:
         st.session_state.agent_pending_action = None