EX MCP Server

""" Core Tool Infrastructure for Zen MCP Tools This module provides the fundamental base class for all tools: - BaseTool: Abstract base class defining the tool interface The BaseTool class defines the core contract that tools must implement and provides common functionality for request validation, error handling, model management, conversation handling, file processing, and response formatting. """ import logging import os from abc import ABC, abstractmethod from typing import TYPE_CHECKING, Any, Optional from mcp.types import TextContent if TYPE_CHECKING: from tools.models import ToolModelCategory from config import MCP_PROMPT_SIZE_LIMIT from src.providers import ModelProvider, ModelProviderRegistry from utils import check_token_limit from utils.conversation_memory import ( ConversationTurn, get_conversation_file_list, get_thread, ) from utils.file_utils import read_file_content, read_files # Import models from tools.models for compatibility try: from tools.models import SPECIAL_STATUS_MODELS, ContinuationOffer, ToolOutput except ImportError: # Fallback in case models haven't been set up yet SPECIAL_STATUS_MODELS = {} ContinuationOffer = None ToolOutput = None logger = logging.getLogger(__name__) class BaseTool(ABC): """ Abstract base class for all Zen MCP tools. This class defines the interface that all tools must implement and provides common functionality for request handling, model creation, and response formatting. CONVERSATION-AWARE FILE PROCESSING: This base class implements the sophisticated dual prioritization strategy for conversation-aware file handling across all tools: 1. FILE DEDUPLICATION WITH NEWEST-FIRST PRIORITY: - When same file appears in multiple conversation turns, newest reference wins - Prevents redundant file embedding while preserving most recent file state - Cross-tool file tracking ensures consistent behavior across analyze → codereview → debug 2. CONVERSATION CONTEXT INTEGRATION: - All tools receive enhanced prompts with conversation history via reconstruct_thread_context() - File references from previous turns are preserved and accessible - Cross-tool knowledge transfer maintains full context without manual file re-specification 3. TOKEN-AWARE FILE EMBEDDING: - Respects model-specific token allocation budgets from ModelContext - Prioritizes conversation history, then newest files, then remaining content - Graceful degradation when token limits are approached 4. STATELESS-TO-STATEFUL BRIDGING: - Tools operate on stateless MCP requests but access full conversation state - Conversation memory automatically injected via continuation_id parameter - Enables natural AI-to-AI collaboration across tool boundaries To create a new tool: 1. Create a new class that inherits from BaseTool 2. Implement all abstract methods 3. Define a request model that inherits from ToolRequest 4. Register the tool in server.py's TOOLS dictionary """ # Class-level cache for OpenRouter registry to avoid multiple loads _openrouter_registry_cache = None @classmethod def _get_openrouter_registry(cls): """Get cached OpenRouter registry instance, creating if needed.""" # Use BaseTool class directly to ensure cache is shared across all subclasses if BaseTool._openrouter_registry_cache is None: from src.providers.openrouter_registry import OpenRouterModelRegistry BaseTool._openrouter_registry_cache = OpenRouterModelRegistry() logger.debug("Created cached OpenRouter registry instance") return BaseTool._openrouter_registry_cache def __init__(self): # Cache tool metadata at initialization to avoid repeated calls self.name = self.get_name() self.description = self.get_description() self.default_temperature = self.get_default_temperature() # Tool initialization complete @abstractmethod def get_name(self) -> str: """ Return the unique name identifier for this tool. This name is used by MCP clients to invoke the tool and must be unique across all registered tools. Returns: str: The tool's unique name (e.g., "review_code", "analyze") """ pass @abstractmethod def get_description(self) -> str: """ Return a detailed description of what this tool does. This description is shown to MCP clients (like Claude) to help them understand when and how to use the tool. It should be comprehensive and include trigger phrases. Returns: str: Detailed tool description with usage examples """ pass @abstractmethod def get_input_schema(self) -> dict[str, Any]: """ Return the JSON Schema that defines this tool's parameters. This schema is used by MCP clients to validate inputs before sending requests. It should match the tool's request model. Returns: Dict[str, Any]: JSON Schema object defining required and optional parameters """ pass @abstractmethod def get_system_prompt(self) -> str: """ Return the system prompt that configures the AI model's behavior. This prompt sets the context and instructions for how the model should approach the task. It's prepended to the user's request. Returns: str: System prompt with role definition and instructions """ pass def get_descriptor(self) -> dict[str, Any]: """Return a machine-readable descriptor for this tool (MVP). Non-breaking default that harvests existing metadata so planners/orchestrators can reason about tools without hardcoding. WorkflowTool will override to add workflow-specific semantics. """ # Basic fields name = self.get_name() try: description = self.get_description() except Exception: description = "" try: annotations = self.get_annotations() or {} except Exception: annotations = {} try: schema = self.get_input_schema() except Exception: schema = {} # Some tool variants (e.g., WorkflowTool) may expose get_required_fields. req_fields: list[str] = [] try: rf = getattr(self, "get_required_fields", None) if callable(rf): req_fields = rf() or [] except Exception: req_fields = [] # Optional: model category for cost/latency hints model_category = None try: get_mc = getattr(self, "get_model_category", None) if callable(get_mc): mc = get_mc() model_category = getattr(mc, "value", getattr(mc, "name", str(mc))) except Exception: model_category = None # Optional: whether tool requires a model requires_model = None try: requires_model = bool(self.requires_model()) except Exception: requires_model = None return { "id": name, "display_name": name, "description": description, "category": "core", "type": "simple", "capabilities": [], "supports_workflow": False, "required_fields": req_fields, "input_schema": schema, "annotations": annotations, **({"model_category": model_category} if model_category else {}), **({"requires_model": requires_model} if requires_model is not None else {}), } def get_annotations(self) -> Optional[dict[str, Any]]: """ Return optional annotations for this tool. Annotations provide hints about tool behavior without being security-critical. They help MCP clients make better decisions about tool usage. Returns: Optional[dict]: Dictionary with annotation fields like readOnlyHint, destructiveHint, etc. Returns None if no annotations are needed. """ return None def requires_model(self) -> bool: """ Return whether this tool requires AI model access. Tools that override execute() to do pure data processing (like planner) should return False to skip model resolution at the MCP boundary. Returns: bool: True if tool needs AI model access (default), False for data-only tools """ return True def is_effective_auto_mode(self) -> bool: """ Check if we're in effective auto mode for schema generation. This determines whether the model parameter should be required in the tool schema. Used at initialization time when schemas are generated. Returns: bool: True if model parameter should be required in the schema """ from config import DEFAULT_MODEL from src.providers.registry import ModelProviderRegistry # Hidden router: if enabled and DEFAULT_MODEL is a sentinel, do NOT require model in schema try: if os.getenv("HIDDEN_MODEL_ROUTER_ENABLED", "true").strip().lower() == "true": sentinels = { s.strip().lower() for s in os.getenv("ROUTER_SENTINEL_MODELS", "glm-4.5-flash,auto").split(",") if s.strip() } if DEFAULT_MODEL.strip().lower() in sentinels: return False except Exception: pass # Case 1: Explicit auto mode if DEFAULT_MODEL.lower() == "auto": return True # Case 2: Model not available (fallback to auto mode) if DEFAULT_MODEL.lower() != "auto": provider = ModelProviderRegistry.get_provider_for_model(DEFAULT_MODEL) if not provider: # If hidden router is enabled, do not require model selection in schema if os.getenv("HIDDEN_MODEL_ROUTER_ENABLED", "true").strip().lower() == "true": return False return True return False def _should_require_model_selection(self, model_name: str) -> bool: """ Check if we should require Claude to select a model at runtime. This is called during request execution to determine if we need to return an error asking Claude to provide a model parameter. Args: model_name: The model name from the request or DEFAULT_MODEL Returns: bool: True if we should require model selection """ # Case 1: Model is explicitly "auto" # Hidden router sentinel should not force runtime selection if os.getenv("HIDDEN_MODEL_ROUTER_ENABLED", "true").strip().lower() == "true": sentinels = {s.strip().lower() for s in os.getenv("ROUTER_SENTINEL_MODELS", "glm-4.5-flash,auto").split(",") if s.strip()} if model_name.strip().lower() in sentinels: return False if model_name.lower() == "auto": return True # Case 2: Requested model is not available from src.providers.registry import ModelProviderRegistry provider = ModelProviderRegistry.get_provider_for_model(model_name) if not provider: logger = logging.getLogger(f"tools.{self.name}") logger.warning(f"Model '{model_name}' is not available with current API keys. Requiring model selection.") return True return False def _get_available_models(self) -> list[str]: """ Get list of models available from enabled providers. Only returns models from providers that have valid API keys configured. This fixes the namespace collision bug where models from disabled providers were shown to Claude, causing routing conflicts. Returns: List of model names from enabled providers only """ from src.providers.registry import ModelProviderRegistry # Get models from enabled providers only (those with valid API keys) all_models = ModelProviderRegistry.get_available_model_names() # Add OpenRouter models if OpenRouter is configured openrouter_key = os.getenv("OPENROUTER_API_KEY") if openrouter_key and openrouter_key != "your_openrouter_api_key_here": try: registry = self._get_openrouter_registry() # Add all aliases from the registry (includes OpenRouter cloud models) for alias in registry.list_aliases(): if alias not in all_models: all_models.append(alias) except Exception as e: import logging logging.debug(f"Failed to add OpenRouter models to enum: {e}") # Add custom models if custom API is configured custom_url = os.getenv("CUSTOM_API_URL") if custom_url: try: registry = self._get_openrouter_registry() # Find all custom models (is_custom=true) for alias in registry.list_aliases(): config = registry.resolve(alias) # Check if this is a custom model that requires custom endpoints if config and config.is_custom: if alias not in all_models: all_models.append(alias) except Exception as e: import logging logging.debug(f"Failed to add custom models to enum: {e}") # Remove duplicates while preserving order seen = set() unique_models = [] for model in all_models: if model not in seen: seen.add(model) unique_models.append(model) return unique_models def get_model_field_schema(self) -> dict[str, Any]: """ Generate the model field schema based on auto mode configuration. When auto mode is enabled, the model parameter becomes required and includes detailed descriptions of each model's capabilities. Returns: Dict containing the model field JSON schema """ import os from config import DEFAULT_MODEL # Check if OpenRouter is configured has_openrouter = bool( os.getenv("OPENROUTER_API_KEY") and os.getenv("OPENROUTER_API_KEY") != "your_openrouter_api_key_here" ) # Use the centralized effective auto mode check if self.is_effective_auto_mode(): # In auto mode, model is required and we provide detailed descriptions model_desc_parts = [ "IMPORTANT: Use the model specified by the user if provided, OR select the most suitable model " "for this specific task based on the requirements and capabilities listed below:" ] # Get descriptions from enabled providers from src.providers.base import ProviderType from src.providers.registry import ModelProviderRegistry # Map provider types to readable names provider_names = { ProviderType.GOOGLE: "Gemini models", ProviderType.OPENAI: "OpenAI models", ProviderType.XAI: "X.AI GROK models", ProviderType.DIAL: "DIAL models", ProviderType.CUSTOM: "Custom models", ProviderType.OPENROUTER: "OpenRouter models", } # Check available providers and add their model descriptions # Start with native providers for provider_type in [ProviderType.GOOGLE, ProviderType.OPENAI, ProviderType.XAI, ProviderType.DIAL]: # Only if this is registered / available provider = ModelProviderRegistry.get_provider(provider_type) if provider: provider_section_added = False for model_name in provider.list_models(respect_restrictions=True): try: # Get model config to extract description model_config = provider.SUPPORTED_MODELS.get(model_name) if model_config and model_config.description: if not provider_section_added: model_desc_parts.append( f"\n{provider_names[provider_type]} - Available when {provider_type.value.upper()}_API_KEY is configured:" ) provider_section_added = True model_desc_parts.append(f"- '{model_name}': {model_config.description}") except Exception: # Skip models without descriptions continue # Add custom models if custom API is configured custom_url = os.getenv("CUSTOM_API_URL") if custom_url: # Load custom models from registry try: registry = self._get_openrouter_registry() model_desc_parts.append(f"\nCustom models via {custom_url}:") # Find all custom models (is_custom=true) for alias in registry.list_aliases(): config = registry.resolve(alias) # Check if this is a custom model that requires custom endpoints if config and config.is_custom: # Format context window context_tokens = config.context_window if context_tokens >= 1_000_000: context_str = f"{context_tokens // 1_000_000}M" elif context_tokens >= 1_000: context_str = f"{context_tokens // 1_000}K" else: context_str = str(context_tokens) desc_line = f"- '{alias}' ({context_str} context): {config.description}" if desc_line not in model_desc_parts: # Avoid duplicates model_desc_parts.append(desc_line) except Exception as e: import logging logging.debug(f"Failed to load custom model descriptions: {e}") model_desc_parts.append(f"\nCustom models: Models available via {custom_url}") if has_openrouter: # Add OpenRouter models with descriptions try: import logging registry = self._get_openrouter_registry() # Group models by their model_name to avoid duplicates seen_models = set() model_configs = [] for alias in registry.list_aliases(): config = registry.resolve(alias) if config and config.model_name not in seen_models: seen_models.add(config.model_name) model_configs.append((alias, config)) # Sort by context window (descending) then by alias model_configs.sort(key=lambda x: (-x[1].context_window, x[0])) if model_configs: model_desc_parts.append("\nOpenRouter models (use these aliases):") for alias, config in model_configs: # Show ALL models so Claude can choose # Format context window in human-readable form context_tokens = config.context_window if context_tokens >= 1_000_000: context_str = f"{context_tokens // 1_000_000}M" elif context_tokens >= 1_000: context_str = f"{context_tokens // 1_000}K" else: context_str = str(context_tokens) # Build description line if config.description: desc = f"- '{alias}' ({context_str} context): {config.description}" else: # Fallback to showing the model name if no description desc = f"- '{alias}' ({context_str} context): {config.model_name}" model_desc_parts.append(desc) # Show all models - no truncation needed except Exception as e: # Log for debugging but don't fail import logging logging.debug(f"Failed to load OpenRouter model descriptions: {e}") # Fallback to simple message model_desc_parts.append( "\nOpenRouter models: If configured, you can also use ANY model available on OpenRouter." ) # Get all available models for the enum (canonicalized) all_models = self._get_available_models() # Canonicalize names by resolving aliases to provider base names def _canonicalize(models: list[str]) -> list[str]: from src.providers.registry import ModelProviderRegistry canon_set = [] seen = set() for m in models: cm = m try: provider = ModelProviderRegistry.get_provider_for_model(m) if provider: caps = provider.get_capabilities(m) cm = caps.model_name or m except Exception: cm = m # Normalize alt uppercase canonical to lowercase variant if cm == "GLM-4": cm = "glm-4" # Filter obvious provider ids low = cm.lower() if low in {"z-ai", "zhipu"}: continue if any(s in low for s in ["-250414", "-airx"]): continue if cm not in seen: seen.add(cm) canon_set.append(cm) return canon_set all_models = _canonicalize(all_models) # Allow 'auto' as a valid option for strict clients; server resolves it if "auto" not in all_models: all_models = ["auto"] + all_models return { "type": "string", "description": "\n".join(model_desc_parts), "enum": all_models, } else: # Normal mode - model is optional with default available_models = self._get_available_models() # Canonicalize + filter questionable names from being exposed in schemas def _canonicalize_and_filter(models: list[str]) -> list[str]: from src.providers.registry import ModelProviderRegistry disallow_exact = {"z-ai", "zhipu"} disallow_substrings = ["-250414", "-airx", "-preview", "-0711", "-0905"] seen = set() cleaned: list[str] = [] for m in models: ml = (m or "").strip() if not ml: continue # Resolve alias to canonical base model when possible cm = ml try: provider = ModelProviderRegistry.get_provider_for_model(ml) if provider: caps = provider.get_capabilities(ml) cm = caps.model_name or ml except Exception: cm = ml # Normalize canonical variations if cm == "GLM-4": cm = "glm-4" low = cm.lower() if low in disallow_exact: continue if any(s in low for s in disallow_substrings): continue if cm not in seen: seen.add(cm) cleaned.append(cm) return cleaned canon_models = _canonicalize_and_filter(available_models) if "auto" not in canon_models: canon_models = ["auto"] + canon_models models_str = ", ".join(f"'{m}'" for m in canon_models) description = f"Model to use. Native models: {models_str}. Use 'auto' to let the server select the best model." if has_openrouter: # Add OpenRouter aliases try: registry = self._get_openrouter_registry() aliases = registry.list_aliases() # Show ALL aliases from the configuration if aliases: # Show all aliases so clients know options all_aliases = sorted(aliases) alias_list = ", ".join(f"'{a}'" for a in all_aliases) description += f" OpenRouter aliases: {alias_list}." else: description += " OpenRouter: Any model available on openrouter.ai." except Exception: description += ( " OpenRouter: Any model available on openrouter.ai " "(e.g., 'gpt-4', 'claude-4-opus', 'mistral-large')." ) description += f" Defaults to '{DEFAULT_MODEL}' if not specified." return { "type": "string", "description": description, } def get_default_temperature(self) -> float: """ Return the default temperature setting for this tool. Override this method to set tool-specific temperature defaults. Lower values (0.0-0.3) for analytical tasks, higher (0.7-1.0) for creative tasks. Returns: float: Default temperature between 0.0 and 1.0 """ return 0.5 def wants_line_numbers_by_default(self) -> bool: """ Return whether this tool wants line numbers added to code files by default. By default, ALL tools get line numbers for precise code references. Line numbers are essential for accurate communication about code locations. Returns: bool: True if line numbers should be added by default for this tool """ return True # All tools get line numbers by default for consistency def get_default_thinking_mode(self) -> str: """ Return the default thinking mode for this tool. Thinking mode controls computational budget for reasoning. Override for tools that need more or less reasoning depth. Returns: str: One of "minimal", "low", "medium", "high", "max" """ return "medium" # Default to medium thinking for better reasoning def get_model_category(self) -> "ToolModelCategory": """ Return the model category for this tool. Model category influences which model is selected in auto mode. Override to specify whether your tool needs extended reasoning, fast response, or balanced capabilities. Returns: ToolModelCategory: Category that influences model selection """ from tools.models import ToolModelCategory return ToolModelCategory.BALANCED @abstractmethod def get_request_model(self): """ Return the Pydantic model class used for validating requests. This model should inherit from ToolRequest and define all parameters specific to this tool. Returns: Type[ToolRequest]: The request model class """ pass def validate_file_paths(self, request) -> Optional[str]: """ Validate that all file paths in the request are absolute. If EX_ALLOW_RELATIVE_PATHS=true, resolve relative paths against the project root (current working directory) and ensure they do not escape that root. This preserves security while improving UX for clients like Claude. Args: request: The validated request object Returns: Optional[str]: Error message if validation fails, None if all paths are valid """ # Only validate files/paths if they exist in the request file_fields = [ "files", "file", "path", "directory", "notebooks", "test_examples", "style_guide_examples", "files_checked", "relevant_files", ] allow_relative = os.getenv("EX_ALLOW_RELATIVE_PATHS", "false").strip().lower() == "true" project_root = os.path.abspath(os.getcwd()) for field_name in file_fields: if hasattr(request, field_name): field_value = getattr(request, field_name) if field_value is None: continue # Handle both single paths and lists of paths is_list = isinstance(field_value, list) paths_to_check = field_value if is_list else [field_value] normalized: list[str] = [] for path in paths_to_check: if not path: normalized.append(path) continue if os.path.isabs(path): # Already absolute - keep as is normalized.append(path) else: if not allow_relative: return f"All file paths must be FULL absolute paths. Invalid path: '{path}'" # Resolve relative path against project root and ensure containment abs_p = os.path.abspath(os.path.join(project_root, path)) if not abs_p.startswith(project_root): return f"Relative path escapes project root: '{path}'" normalized.append(abs_p) # Write back normalized values to the request model for downstream consumers try: setattr(request, field_name, normalized if is_list else (normalized[0] if normalized else field_value)) except Exception: # If we cannot set, we still validated; but prefer not to fail here pass return None def _validate_token_limit(self, content: str, content_type: str = "Content") -> None: """ Validate that content doesn't exceed the MCP prompt size limit. Args: content: The content to validate content_type: Description of the content type for error messages Raises: ValueError: If content exceeds size limit """ is_valid, token_count = check_token_limit(content, MCP_PROMPT_SIZE_LIMIT) if not is_valid: error_msg = f"~{token_count:,} tokens. Maximum is {MCP_PROMPT_SIZE_LIMIT:,} tokens." logger.error(f"{self.name} tool {content_type.lower()} validation failed: {error_msg}") raise ValueError(f"{content_type} too large: {error_msg}") logger.debug(f"{self.name} tool {content_type.lower()} token validation passed: {token_count:,} tokens") def get_model_provider(self, model_name: str) -> ModelProvider: """ Get the appropriate model provider for the given model name. This method performs runtime validation to ensure the requested model is actually available with the current API key configuration. Args: model_name: Name of the model to get provider for Returns: ModelProvider: The provider instance for the model Raises: ValueError: If the model is not available or provider not found """ try: provider = ModelProviderRegistry.get_provider_for_model(model_name) if not provider: logger.error(f"No provider found for model '{model_name}' in {self.name} tool") available_models = ModelProviderRegistry.get_available_models() raise ValueError(f"Model '{model_name}' is not available. Available models: {available_models}") return provider except Exception as e: logger.error(f"Failed to get provider for model '{model_name}' in {self.name} tool: {e}") raise # === CONVERSATION AND FILE HANDLING METHODS === def get_conversation_embedded_files(self, continuation_id: Optional[str]) -> list[str]: """ Get list of files already embedded in conversation history. This method returns the list of files that have already been embedded in the conversation history for a given continuation thread. Tools can use this to avoid re-embedding files that are already available in the conversation context. Args: continuation_id: Thread continuation ID, or None for new conversations Returns: list[str]: List of file paths already embedded in conversation history """ if not continuation_id: # New conversation, no files embedded yet return [] thread_context = get_thread(continuation_id) if not thread_context: # Thread not found, no files embedded return [] embedded_files = get_conversation_file_list(thread_context) logger.debug(f"[FILES] {self.name}: Found {len(embedded_files)} embedded files") return embedded_files def filter_new_files(self, requested_files: list[str], continuation_id: Optional[str]) -> list[str]: """ Filter out files that are already embedded in conversation history. This method prevents duplicate file embeddings by filtering out files that have already been embedded in the conversation history. This optimizes token usage while ensuring tools still have logical access to all requested files through conversation history references. Args: requested_files: List of files requested for current tool execution continuation_id: Thread continuation ID, or None for new conversations Returns: list[str]: List of files that need to be embedded (not already in history) """ logger.debug(f"[FILES] {self.name}: Filtering {len(requested_files)} requested files") if not continuation_id: # New conversation, all files are new logger.debug(f"[FILES] {self.name}: New conversation, all {len(requested_files)} files are new") return requested_files try: embedded_files = set(self.get_conversation_embedded_files(continuation_id)) logger.debug(f"[FILES] {self.name}: Found {len(embedded_files)} embedded files in conversation") # Safety check: If no files are marked as embedded but we have a continuation_id, # this might indicate an issue with conversation history. Be conservative. if not embedded_files: logger.debug(f"{self.name} tool: No files found in conversation history for thread {continuation_id}") logger.debug( f"[FILES] {self.name}: No embedded files found, returning all {len(requested_files)} requested files" ) return requested_files # Return only files that haven't been embedded yet new_files = [f for f in requested_files if f not in embedded_files] logger.debug( f"[FILES] {self.name}: After filtering: {len(new_files)} new files, {len(requested_files) - len(new_files)} already embedded" ) logger.debug(f"[FILES] {self.name}: New files to embed: {new_files}") # Log filtering results for debugging if len(new_files) < len(requested_files): skipped = [f for f in requested_files if f in embedded_files] logger.debug( f"{self.name} tool: Filtering {len(skipped)} files already in conversation history: {', '.join(skipped)}" ) logger.debug(f"[FILES] {self.name}: Skipped (already embedded): {skipped}") return new_files except Exception as e: # If there's any issue with conversation history lookup, be conservative # and include all files rather than risk losing access to needed files logger.warning(f"{self.name} tool: Error checking conversation history for {continuation_id}: {e}") logger.warning(f"{self.name} tool: Including all requested files as fallback") logger.debug( f"[FILES] {self.name}: Exception in filter_new_files, returning all {len(requested_files)} files as fallback" ) return requested_files def format_conversation_turn(self, turn: ConversationTurn) -> list[str]: """ Format a conversation turn for display in conversation history. Tools can override this to provide custom formatting for their responses while maintaining the standard structure for cross-tool compatibility. This method is called by build_conversation_history when reconstructing conversation context, allowing each tool to control how its responses appear in subsequent conversation turns. Args: turn: The conversation turn to format (from utils.conversation_memory) Returns: list[str]: Lines of formatted content for this turn Example: Default implementation returns: ["Files used in this turn: file1.py, file2.py", "", "Response content..."] Tools can override to add custom sections, formatting, or metadata display. """ parts = [] # Add files context if present if turn.files: parts.append(f"Files used in this turn: {', '.join(turn.files)}") parts.append("") # Empty line for readability # Add the actual content parts.append(turn.content) return parts def handle_prompt_file(self, files: Optional[list[str]]) -> tuple[Optional[str], Optional[list[str]]]: """ Check for and handle prompt.txt in the files list. If prompt.txt is found, reads its content and removes it from the files list. This file is treated specially as the main prompt, not as an embedded file. This mechanism allows us to work around MCP's ~25K token limit by having Claude save large prompts to a file, effectively using the file transfer mechanism to bypass token constraints while preserving response capacity. Args: files: List of file paths (will be translated for current environment) Returns: tuple: (prompt_content, updated_files_list) """ if not files: return None, files prompt_content = None updated_files = [] for file_path in files: # Check if the filename is exactly "prompt.txt" # This ensures we don't match files like "myprompt.txt" or "prompt.txt.bak" if os.path.basename(file_path) == "prompt.txt": try: # Read prompt.txt content and extract just the text content, _ = read_file_content(file_path) # Extract the content between the file markers if "--- BEGIN FILE:" in content and "--- END FILE:" in content: lines = content.split("\n") in_content = False content_lines = [] for line in lines: if line.startswith("--- BEGIN FILE:"): in_content = True continue elif line.startswith("--- END FILE:"): break elif in_content: content_lines.append(line) prompt_content = "\n".join(content_lines) else: # Fallback: if it's already raw content (from tests or direct input) # and doesn't have error markers, use it directly if not content.startswith("\n--- ERROR"): prompt_content = content else: prompt_content = None except Exception: # If we can't read the file, we'll just skip it # The error will be handled elsewhere pass else: # Keep the original path in the files list (will be translated later by read_files) updated_files.append(file_path) return prompt_content, updated_files if updated_files else None def get_prompt_content_for_size_validation(self, user_content: str) -> str: """ Get the content that should be validated for MCP prompt size limits. This hook method allows tools to specify what content should be checked against the MCP transport size limit. By default, it returns the user content, but can be overridden to exclude conversation history when needed. Args: user_content: The user content that would normally be validated Returns: The content that should actually be validated for size limits """ # Default implementation: validate the full user content return user_content def check_prompt_size(self, text: str) -> Optional[dict[str, Any]]: """ Check if USER INPUT text is too large for MCP transport boundary. IMPORTANT: This method should ONLY be used to validate user input that crosses the Claude CLI ↔ MCP Server transport boundary. It should NOT be used to limit internal MCP Server operations. Args: text: The user input text to check (NOT internal prompt content) Returns: Optional[Dict[str, Any]]: Response asking for file handling if too large, None otherwise """ if text and len(text) > MCP_PROMPT_SIZE_LIMIT: return { "status": "resend_prompt", "content": ( f"MANDATORY ACTION REQUIRED: The prompt is too large for MCP's token limits (>{MCP_PROMPT_SIZE_LIMIT:,} characters). " "YOU MUST IMMEDIATELY save the prompt text to a temporary file named 'prompt.txt' in the working directory. " "DO NOT attempt to shorten or modify the prompt. SAVE IT AS-IS to 'prompt.txt'. " "Then resend the request with the absolute file path to 'prompt.txt' in the files parameter (must be FULL absolute path - DO NOT SHORTEN), " "along with any other files you wish to share as context. Leave the prompt text itself empty or very brief in the new request. " "This is the ONLY way to handle large prompts - you MUST follow these exact steps." ), "content_type": "text", "metadata": { "prompt_size": len(text), "limit": MCP_PROMPT_SIZE_LIMIT, "instructions": "MANDATORY: Save prompt to 'prompt.txt' in current folder and include absolute path in files parameter. DO NOT modify or shorten the prompt.", }, } return None def _prepare_file_content_for_prompt( self, request_files: list[str], continuation_id: Optional[str], context_description: str = "New files", max_tokens: Optional[int] = None, reserve_tokens: int = 1_000, remaining_budget: Optional[int] = None, arguments: Optional[dict] = None, model_context: Optional[Any] = None, ) -> tuple[str, list[str]]: """ Centralized file processing implementing dual prioritization strategy. This method is the heart of conversation-aware file processing across all tools. Args: request_files: List of files requested for current tool execution continuation_id: Thread continuation ID, or None for new conversations context_description: Description for token limit validation (e.g. "Code", "New files") max_tokens: Maximum tokens to use (defaults to remaining budget or model-specific content allocation) reserve_tokens: Tokens to reserve for additional prompt content (default 1K) remaining_budget: Remaining token budget after conversation history (from server.py) arguments: Original tool arguments (used to extract _remaining_tokens if available) model_context: Model context object with all model information including token allocation Returns: tuple[str, list[str]]: (formatted_file_content, actually_processed_files) - formatted_file_content: Formatted file content string ready for prompt inclusion - actually_processed_files: List of individual file paths that were actually read and embedded (directories are expanded to individual files) """ if not request_files: return "", [] # Extract remaining budget from arguments if available if remaining_budget is None: # Use provided arguments or fall back to stored arguments from execute() args_to_use = arguments or getattr(self, "_current_arguments", {}) remaining_budget = args_to_use.get("_remaining_tokens") # Use remaining budget if provided, otherwise fall back to max_tokens or model-specific default if remaining_budget is not None: effective_max_tokens = remaining_budget - reserve_tokens elif max_tokens is not None: effective_max_tokens = max_tokens - reserve_tokens else: # Use model_context for token allocation if not model_context: # Try to get from stored attributes as fallback model_context = getattr(self, "_model_context", None) if not model_context: logger.error( f"[FILES] {self.name}: _prepare_file_content_for_prompt called without model_context. " "This indicates an incorrect call sequence in the tool's implementation." ) raise RuntimeError("Model context not provided for file preparation.") # This is now the single source of truth for token allocation. try: token_allocation = model_context.calculate_token_allocation() # Standardize on `file_tokens` for consistency and correctness. effective_max_tokens = token_allocation.file_tokens - reserve_tokens logger.debug( f"[FILES] {self.name}: Using model context for {model_context.model_name}: " f"{token_allocation.file_tokens:,} file tokens from {token_allocation.total_tokens:,} total" ) except Exception as e: logger.error( f"[FILES] {self.name}: Failed to calculate token allocation from model context: {e}", exc_info=True ) # If the context exists but calculation fails, we still need to prevent a crash. # A loud error is logged, and we fall back to a safe default. effective_max_tokens = 100_000 - reserve_tokens # Ensure we have a reasonable minimum budget effective_max_tokens = max(1000, effective_max_tokens) files_to_embed = self.filter_new_files(request_files, continuation_id) logger.debug(f"[FILES] {self.name}: Will embed {len(files_to_embed)} files after filtering") # Log the specific files for debugging/testing if files_to_embed: logger.info( f"[FILE_PROCESSING] {self.name} tool will embed new files: {', '.join([os.path.basename(f) for f in files_to_embed])}" ) else: logger.info( f"[FILE_PROCESSING] {self.name} tool: No new files to embed (all files already in conversation history)" ) content_parts = [] actually_processed_files = [] # Read content of new files only if files_to_embed: logger.debug(f"{self.name} tool embedding {len(files_to_embed)} new files: {', '.join(files_to_embed)}") logger.debug( f"[FILES] {self.name}: Starting file embedding with token budget {effective_max_tokens + reserve_tokens:,}" ) try: # Before calling read_files, expand directories to get individual file paths from utils.file_utils import expand_paths expanded_files = expand_paths(files_to_embed) logger.debug( f"[FILES] {self.name}: Expanded {len(files_to_embed)} paths to {len(expanded_files)} individual files" ) file_content = read_files( files_to_embed, max_tokens=effective_max_tokens + reserve_tokens, reserve_tokens=reserve_tokens, include_line_numbers=self.wants_line_numbers_by_default(), ) self._validate_token_limit(file_content, context_description) content_parts.append(file_content) # Track the expanded files as actually processed actually_processed_files.extend(expanded_files) # Estimate tokens for debug logging from utils.token_utils import estimate_tokens content_tokens = estimate_tokens(file_content) logger.debug( f"{self.name} tool successfully embedded {len(files_to_embed)} files ({content_tokens:,} tokens)" ) logger.debug(f"[FILES] {self.name}: Successfully embedded files - {content_tokens:,} tokens used") logger.debug( f"[FILES] {self.name}: Actually processed {len(actually_processed_files)} individual files" ) except Exception as e: logger.error(f"{self.name} tool failed to embed files {files_to_embed}: {type(e).__name__}: {e}") logger.debug(f"[FILES] {self.name}: File embedding failed - {type(e).__name__}: {e}") raise else: logger.debug(f"[FILES] {self.name}: No files to embed after filtering") # Generate note about files already in conversation history if continuation_id and len(files_to_embed) < len(request_files): embedded_files = self.get_conversation_embedded_files(continuation_id) skipped_files = [f for f in request_files if f in embedded_files] if skipped_files: logger.debug( f"{self.name} tool skipping {len(skipped_files)} files already in conversation history: {', '.join(skipped_files)}" ) logger.debug(f"[FILES] {self.name}: Adding note about {len(skipped_files)} skipped files") if content_parts: content_parts.append("\n\n") note_lines = [ "--- NOTE: Additional files referenced in conversation history ---", "The following files are already available in our conversation context:", "\n".join(f" - {f}" for f in skipped_files), "--- END NOTE ---", ] content_parts.append("\n".join(note_lines)) else: logger.debug(f"[FILES] {self.name}: No skipped files to note") result = "".join(content_parts) if content_parts else "" logger.debug( f"[FILES] {self.name}: _prepare_file_content_for_prompt returning {len(result)} chars, {len(actually_processed_files)} processed files" ) return result, actually_processed_files def get_websearch_instruction(self, use_websearch: bool, tool_specific: Optional[str] = None) -> str: """ Generate standardized web search instruction based on the use_websearch parameter. Args: use_websearch: Whether web search is enabled tool_specific: Optional tool-specific search guidance Returns: str: Web search instruction to append to prompt, or empty string """ if not use_websearch: return "" base_instruction = """ WEB SEARCH CAPABILITY: You can request Claude to perform web searches to enhance your analysis with current information! IMPORTANT: When you identify areas where web searches would significantly improve your response (such as checking current documentation, finding recent solutions, verifying best practices, or gathering community insights), you MUST explicitly instruct Claude to perform specific web searches and then respond back using the continuation_id from this response to continue the analysis. Use clear, direct language based on the value of the search: For valuable supplementary information: "Please perform a web search on '[specific topic/query]' and then continue this analysis using the continuation_id from this response if you find relevant information." For important missing information: "Please search for '[specific topic/query]' and respond back with the findings using the continuation_id from this response - this information is needed to provide a complete analysis." For critical/essential information: "SEARCH REQUIRED: Please immediately perform a web search on '[specific topic/query]' and respond back with the results using the continuation_id from this response. Cannot provide accurate analysis without this current information." This ensures you get the most current and comprehensive information while maintaining conversation context through the continuation_id.""" if tool_specific: return f"""{base_instruction} {tool_specific} When recommending searches, be specific about what information you need and why it would improve your analysis.""" # Default instruction for all tools return f"""{base_instruction} Consider requesting searches for: - Current documentation and API references - Recent best practices and patterns - Known issues and community solutions - Framework updates and compatibility - Security advisories and patches - Performance benchmarks and optimizations When recommending searches, be specific about what information you need and why it would improve your analysis. Always remember to instruct Claude to use the continuation_id from this response when providing search results.""" def get_language_instruction(self) -> str: """ Generate language instruction based on LOCALE configuration. Returns: str: Language instruction to prepend to prompt, or empty string if no locale set """ # Read LOCALE directly from environment to support dynamic changes # This allows tests to modify os.environ["LOCALE"] and see the changes import os locale = os.getenv("LOCALE", "").strip() if not locale: return "" # Simple language instruction return f"Always respond in {locale}.\n\n" # === ABSTRACT METHODS FOR SIMPLE TOOLS === @abstractmethod async def prepare_prompt(self, request) -> str: """ Prepare the complete prompt for the AI model. This method should construct the full prompt by combining: - System prompt from get_system_prompt() - File content from _prepare_file_content_for_prompt() - Conversation history from reconstruct_thread_context() - User's request and any tool-specific context Args: request: The validated request object Returns: str: Complete prompt ready for the AI model """ pass def format_response(self, response: str, request, model_info: dict = None) -> str: """ Format the AI model's response for the user. This method allows tools to post-process the model's response, adding structure, validation, or additional context. The default implementation returns the response unchanged. Tools can override this method to add custom formatting. Args: response: Raw response from the AI model request: The original request object model_info: Optional model information and metadata Returns: str: Formatted response ready for the user """ return response # === IMPLEMENTATION METHODS === # These will be provided in a full implementation but are inherited from current base.py # for now to maintain compatibility. async def execute(self, arguments: dict[str, Any]) -> list[TextContent]: """Execute the tool - will be inherited from existing base.py for now.""" # This will be implemented by importing from the current base.py # for backward compatibility during the migration raise NotImplementedError("Subclasses must implement execute method") def _should_require_model_selection(self, model_name: str) -> bool: """ Check if we should require Claude to select a model at runtime. This is called during request execution to determine if we need to return an error asking Claude to provide a model parameter. Args: model_name: The model name from the request or DEFAULT_MODEL Returns: bool: True if we should require model selection """ # Case 1: Model is explicitly "auto" if model_name.lower() == "auto": return True # Case 2: Requested model is not available from src.providers.registry import ModelProviderRegistry provider = ModelProviderRegistry.get_provider_for_model(model_name) if not provider: logger.warning(f"Model '{model_name}' is not available with current API keys. Requiring model selection.") return True return False def _get_available_models(self) -> list[str]: """ Get list of models available from enabled providers. Only returns models from providers that have valid API keys configured. This fixes the namespace collision bug where models from disabled providers were shown to Claude, causing routing conflicts. Returns: List of model names from enabled providers only """ from src.providers.registry import ModelProviderRegistry # Get models from enabled providers only (those with valid API keys) all_models = ModelProviderRegistry.get_available_model_names() # Add OpenRouter models if OpenRouter is configured openrouter_key = os.getenv("OPENROUTER_API_KEY") if openrouter_key and openrouter_key != "your_openrouter_api_key_here": try: from config import OPENROUTER_MODELS all_models.extend(OPENROUTER_MODELS) except ImportError: pass return sorted(set(all_models)) def _resolve_model_context(self, arguments: dict, request) -> tuple[str, Any]: """ Resolve model context and name using centralized logic. This method extracts the model resolution logic from execute() so it can be reused by tools that override execute() (like debug tool) without duplicating code. Args: arguments: Dictionary of arguments from the MCP client request: The validated request object Returns: tuple[str, ModelContext]: (resolved_model_name, model_context) Raises: ValueError: If model resolution fails or model selection is required """ # MODEL RESOLUTION NOW HAPPENS AT MCP BOUNDARY # Extract pre-resolved model context from server.py model_context = arguments.get("_model_context") resolved_model_name = arguments.get("_resolved_model_name") if model_context and resolved_model_name: # Model was already resolved at MCP boundary model_name = resolved_model_name logger.debug(f"Using pre-resolved model '{model_name}' from MCP boundary") else: # Fallback for direct execute calls model_name = getattr(request, "model", None) if not model_name: from config import DEFAULT_MODEL model_name = DEFAULT_MODEL logger.debug(f"Using fallback model resolution for '{model_name}' (test mode)") # Hidden router: resolve to a real model if sentinel if os.getenv("HIDDEN_MODEL_ROUTER_ENABLED", "true").strip().lower() == "true": sentinels = {s.strip().lower() for s in os.getenv("ROUTER_SENTINEL_MODELS", "glm-4.5-flash,auto").split(",") if s.strip()} if model_name.strip().lower() in sentinels: from src.router.service import RouterService routed = RouterService().choose_model(model_name) if routed and getattr(routed, "chosen", None): logger.debug(f"Hidden-router selected '{routed.chosen}' reason='{routed.reason}'") model_name = routed.chosen # For tests: Check if we should require model selection if self._should_require_model_selection(model_name): from src.providers.registry import ModelProviderRegistry tool_category = self.get_model_category() suggested_model = ModelProviderRegistry.get_preferred_fallback_model(tool_category) available_models = self._get_available_models() error_message = ( f"Model '{model_name}' is not available with current API keys. " f"Available models: {', '.join(available_models)}. " f"Suggested model for {self.get_name()}: '{suggested_model}' " f"(category: {tool_category.value})" ) raise ValueError(error_message) # Create model context for tests from utils.model_context import ModelContext model_context = ModelContext(model_name) return model_name, model_context def validate_and_correct_temperature(self, temperature: float, model_context: Any) -> tuple[float, list[str]]: """ Validate and correct temperature for the specified model. This method ensures that the temperature value is within the valid range for the specific model being used. Different models have different temperature constraints (e.g., o1 models require temperature=1.0, GPT models support 0-2). Args: temperature: Temperature value to validate model_context: Model context object containing model name, provider, and capabilities Returns: Tuple of (corrected_temperature, warning_messages) """ try: # Use model context capabilities directly - clean OOP approach capabilities = model_context.capabilities constraint = capabilities.temperature_constraint warnings = [] if not constraint.validate(temperature): corrected = constraint.get_corrected_value(temperature) warning = ( f"Temperature {temperature} invalid for {model_context.model_name}. " f"{constraint.get_description()}. Using {corrected} instead." ) warnings.append(warning) return corrected, warnings return temperature, warnings except Exception as e: # If validation fails for any reason, use the original temperature # and log a warning (but don't fail the request) logger.warning(f"Temperature validation failed for {model_context.model_name}: {e}") return temperature, [f"Temperature validation failed: {e}"] def _validate_image_limits( self, images: Optional[list[str]], model_context: Optional[Any] = None, continuation_id: Optional[str] = None ) -> Optional[dict]: """ Validate image size and count against model capabilities. This performs strict validation to ensure we don't exceed model-specific image limits. Uses capability-based validation with actual model configuration rather than hard-coded limits. Args: images: List of image paths/data URLs to validate model_context: Model context object containing model name, provider, and capabilities continuation_id: Optional continuation ID for conversation context Returns: Optional[dict]: Error response if validation fails, None if valid """ if not images: return None # Import here to avoid circular imports import base64 from pathlib import Path # Handle legacy calls (positional model_name string) if isinstance(model_context, str): # Legacy call: _validate_image_limits(images, "model-name") logger.warning( "Legacy _validate_image_limits call with model_name string. Use model_context object instead." ) try: from utils.model_context import ModelContext model_context = ModelContext(model_context) except Exception as e: logger.warning(f"Failed to create model context from legacy model_name: {e}") # Generic error response for any unavailable model return { "status": "error", "content": f"Model '{model_context}' is not available. {str(e)}", "content_type": "text", "metadata": { "error_type": "validation_error", "model_name": model_context, "supports_images": None, # Unknown since model doesn't exist "image_count": len(images) if images else 0, }, } if not model_context: # Get from tool's stored context as fallback model_context = getattr(self, "_model_context", None) if not model_context: logger.warning("No model context available for image validation") return None try: # Use model context capabilities directly - clean OOP approach capabilities = model_context.capabilities model_name = model_context.model_name except Exception as e: logger.warning(f"Failed to get capabilities from model_context for image validation: {e}") # Generic error response when capabilities cannot be accessed model_name = getattr(model_context, "model_name", "unknown") return { "status": "error", "content": f"Model '{model_name}' is not available. {str(e)}", "content_type": "text", "metadata": { "error_type": "validation_error", "model_name": model_name, "supports_images": None, # Unknown since model capabilities unavailable "image_count": len(images) if images else 0, }, } # Check if model supports images if not capabilities.supports_images: return { "status": "error", "content": ( f"Image support not available: Model '{model_name}' does not support image processing. " f"Please use a vision-capable model such as 'gemini-2.5-flash', 'o3', " f"or 'claude-3-opus' for image analysis tasks." ), "content_type": "text", "metadata": { "error_type": "validation_error", "model_name": model_name, "supports_images": False, "image_count": len(images), }, } # Get model image limits from capabilities max_images = 5 # Default max number of images max_size_mb = capabilities.max_image_size_mb # Check image count if len(images) > max_images: return { "status": "error", "content": ( f"Too many images: Model '{model_name}' supports a maximum of {max_images} images, " f"but {len(images)} were provided. Please reduce the number of images." ), "content_type": "text", "metadata": { "error_type": "validation_error", "model_name": model_name, "image_count": len(images), "max_images": max_images, }, } # Calculate total size of all images total_size_mb = 0.0 for image_path in images: try: if image_path.startswith("data:image/"): # Handle data URL: data:image/png;base64,iVBORw0... _, data = image_path.split(",", 1) # Base64 encoding increases size by ~33%, so decode to get actual size actual_size = len(base64.b64decode(data)) total_size_mb += actual_size / (1024 * 1024) else: # Handle file path path = Path(image_path) if path.exists(): file_size = path.stat().st_size total_size_mb += file_size / (1024 * 1024) else: logger.warning(f"Image file not found: {image_path}") # Assume a reasonable size for missing files to avoid breaking validation total_size_mb += 1.0 # 1MB assumption except Exception as e: logger.warning(f"Failed to get size for image {image_path}: {e}") # Assume a reasonable size for problematic files total_size_mb += 1.0 # 1MB assumption # Apply 40MB cap for custom models if needed effective_limit_mb = max_size_mb try: from src.providers.base import ProviderType # ModelCapabilities dataclass has provider field defined if capabilities.provider == ProviderType.CUSTOM: effective_limit_mb = min(max_size_mb, 40.0) except Exception: pass # Validate against size limit if total_size_mb > effective_limit_mb: return { "status": "error", "content": ( f"Image size limit exceeded: Model '{model_name}' supports maximum {effective_limit_mb:.1f}MB " f"for all images combined, but {total_size_mb:.1f}MB was provided. " f"Please reduce image sizes or count and try again." ), "content_type": "text", "metadata": { "error_type": "validation_error", "model_name": model_name, "total_size_mb": round(total_size_mb, 2), "limit_mb": round(effective_limit_mb, 2), "image_count": len(images), "supports_images": True, }, } # All validations passed logger.debug(f"Image validation passed: {len(images)} images, {total_size_mb:.1f}MB total") return None def _parse_response(self, raw_text: str, request, model_info: Optional[dict] = None): """Parse response - will be inherited for now.""" # Implementation inherited from current base.py raise NotImplementedError("Subclasses must implement _parse_response method")