massgen 0.1.0a3__py3-none-any.whl → 0.1.2__py3-none-any.whl

massgen/formatter/_gemini_formatter.py

@@ -0,0 +1,448 @@
+# -*- coding: utf-8 -*-
+"""
+Gemini formatter for message formatting, coordination prompts, and structured output parsing.
+"""
+
+import json
+import logging
+import re
+from typing import Any, Dict, List, Optional
+
+from ._formatter_base import FormatterBase
+
+logger = logging.getLogger(__name__)
+
+
+class GeminiFormatter(FormatterBase):
+    def format_messages(self, messages: List[Dict[str, Any]]) -> str:
+        """
+        Build conversation content string from message history.
+
+        Behavior mirrors the formatting used previously in the Gemini backend:
+        - System messages buffered separately then prepended.
+        - User => "User: {content}"
+        - Assistant => "Assistant: {content}"
+        - Tool => "Tool Result: {content}"
+        """
+        conversation_content = ""
+        system_message = ""
+
+        for msg in messages:
+            role = msg.get("role")
+            if role == "system":
+                system_message = msg.get("content", "")
+            elif role == "user":
+                conversation_content += f"User: {msg.get('content', '')}\n"
+            elif role == "assistant":
+                conversation_content += f"Assistant: {msg.get('content', '')}\n"
+            elif role == "tool":
+                tool_output = msg.get("content", "")
+                conversation_content += f"Tool Result: {tool_output}\n"
+
+        # Combine system message and conversation
+        full_content = ""
+        if system_message:
+            full_content += f"{system_message}\n\n"
+        full_content += conversation_content
+        return full_content
+
+    def format_tools(self, tools: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+        """
+        Gemini uses SDK-native tool format, not reformatting.
+        """
+        return tools or []
+
+    def format_mcp_tools(self, mcp_functions: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """
+        MCP tools are passed via SDK sessions in stream_with_tools; not function declarations.
+        """
+        return []
+
+    # Coordination helpers
+
+    def has_coordination_tools(self, tools: List[Dict[str, Any]]) -> bool:
+        """Detect if tools contain vote/new_answer coordination tools."""
+        if not tools:
+            return False
+
+        tool_names = set()
+        for tool in tools:
+            if tool.get("type") == "function":
+                if "function" in tool:
+                    tool_names.add(tool["function"].get("name", ""))
+                elif "name" in tool:
+                    tool_names.add(tool.get("name", ""))
+
+        return "vote" in tool_names and "new_answer" in tool_names
+
+    def build_structured_output_prompt(self, base_content: str, valid_agent_ids: Optional[List[str]] = None) -> str:
+        """Build prompt that encourages structured output for coordination."""
+        agent_list = ""
+        if valid_agent_ids:
+            agent_list = f"Valid agents: {', '.join(valid_agent_ids)}"
+
+        return f"""{base_content}
+
+IMPORTANT: You must respond with a structured JSON decision at the end of your response.
+
+If you want to VOTE for an existing agent's answer:
+{{
+  "action_type": "vote",
+  "vote_data": {{
+    "action": "vote",
+    "agent_id": "agent1",  // Choose from: {agent_list or "agent1, agent2, agent3, etc."}
+    "reason": "Brief reason for your vote"
+  }}
+}}
+
+If you want to provide a NEW ANSWER:
+{{
+  "action_type": "new_answer",
+  "answer_data": {{
+    "action": "new_answer",
+    "content": "Your complete improved answer here"
+  }}
+}}
+
+Make your decision and include the JSON at the very end of your response."""
+
+    def extract_structured_response(self, response_text: str) -> Optional[Dict[str, Any]]:
+        """Extract structured JSON response from model output."""
+        try:
+            # Strategy 0: Look for JSON inside markdown code blocks first
+            markdown_json_pattern = r"```json\s*(\{.*?\})\s*```"
+            markdown_matches = re.findall(markdown_json_pattern, response_text, re.DOTALL)
+
+            for match in reversed(markdown_matches):
+                try:
+                    parsed = json.loads(match.strip())
+                    if isinstance(parsed, dict) and "action_type" in parsed:
+                        return parsed
+                except json.JSONDecodeError:
+                    continue
+
+            # Strategy 1: Look for complete JSON blocks with proper braces
+            json_pattern = r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}"
+            json_matches = re.findall(json_pattern, response_text, re.DOTALL)
+
+            # Try parsing each match (in reverse order - last one first)
+            for match in reversed(json_matches):
+                try:
+                    cleaned_match = match.strip()
+                    parsed = json.loads(cleaned_match)
+                    if isinstance(parsed, dict) and "action_type" in parsed:
+                        return parsed
+                except json.JSONDecodeError:
+                    continue
+
+            # Strategy 2: Look for JSON blocks with nested braces (more complex)
+            brace_count = 0
+            json_start = -1
+
+            for i, char in enumerate(response_text):
+                if char == "{":
+                    if brace_count == 0:
+                        json_start = i
+                    brace_count += 1
+                elif char == "}":
+                    brace_count -= 1
+                    if brace_count == 0 and json_start >= 0:
+                        # Found a complete JSON block
+                        json_block = response_text[json_start : i + 1]
+                        try:
+                            parsed = json.loads(json_block)
+                            if isinstance(parsed, dict) and "action_type" in parsed:
+                                return parsed
+                        except json.JSONDecodeError:
+                            pass
+                        json_start = -1
+
+            # Strategy 3: Line-by-line approach (fallback)
+            lines = response_text.strip().split("\n")
+            json_candidates = []
+
+            for i, line in enumerate(lines):
+                stripped = line.strip()
+                if stripped.startswith("{") and stripped.endswith("}"):
+                    json_candidates.append(stripped)
+                elif stripped.startswith("{"):
+                    # Multi-line JSON - collect until closing brace
+                    json_text = stripped
+                    for j in range(i + 1, len(lines)):
+                        json_text += "\n" + lines[j].strip()
+                        if lines[j].strip().endswith("}"):
+                            json_candidates.append(json_text)
+                            break
+
+            # Try to parse each candidate
+            for candidate in reversed(json_candidates):
+                try:
+                    parsed = json.loads(candidate)
+                    if isinstance(parsed, dict) and "action_type" in parsed:
+                        return parsed
+                except json.JSONDecodeError:
+                    continue
+
+            return None
+
+        except Exception:
+            return None
+
+    def convert_structured_to_tool_calls(self, structured_response: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """Convert structured response to tool call format."""
+        action_type = structured_response.get("action_type")
+
+        if action_type == "vote":
+            vote_data = structured_response.get("vote_data", {})
+            return [
+                {
+                    "id": f"vote_{abs(hash(str(vote_data))) % 10000 + 1}",
+                    "type": "function",
+                    "function": {
+                        "name": "vote",
+                        "arguments": {
+                            "agent_id": vote_data.get("agent_id", ""),
+                            "reason": vote_data.get("reason", ""),
+                        },
+                    },
+                },
+            ]
+
+        elif action_type == "new_answer":
+            answer_data = structured_response.get("answer_data", {})
+            return [
+                {
+                    "id": f"new_answer_{abs(hash(str(answer_data))) % 10000 + 1}",
+                    "type": "function",
+                    "function": {
+                        "name": "new_answer",
+                        "arguments": {"content": answer_data.get("content", "")},
+                    },
+                },
+            ]
+
+        return []
+
+    # Custom tools formatting for Gemini
+
+    def format_custom_tools(
+        self,
+        custom_tools: List[Dict[str, Any]],
+        return_sdk_objects: bool = True,
+    ) -> List[Any]:
+        """
+        Convert custom tools from OpenAI Chat Completions format to Gemini format.
+
+        Can return either SDK FunctionDeclaration objects (default) or Gemini-format dictionaries.
+
+        OpenAI format:
+            [{"type": "function", "function": {"name": ..., "description": ..., "parameters": {...}}}]
+
+        Gemini dictionary format:
+            [{"name": ..., "description": ..., "parameters": {...}}]
+
+        Gemini SDK format:
+            [FunctionDeclaration(name=..., description=..., parameters=Schema(...))]
+
+        Args:
+            custom_tools: List of tools in OpenAI Chat Completions format
+            return_sdk_objects: If True, return FunctionDeclaration objects;
+                               if False, return Gemini-format dictionaries
+
+        Returns:
+            List of tools in Gemini SDK format (default) or dictionary format
+        """
+        if not custom_tools:
+            return []
+
+        # Step 1: Convert to Gemini dictionary format
+        gemini_dicts = self._convert_to_gemini_dict_format(custom_tools)
+
+        if not return_sdk_objects:
+            return gemini_dicts
+
+        # Step 2: Convert dictionaries to SDK FunctionDeclaration objects
+        return self._convert_to_function_declarations(gemini_dicts)
+
+    def _convert_to_gemini_dict_format(self, custom_tools: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+        """
+        Convert OpenAI format to Gemini dictionary format (intermediate step).
+
+        Args:
+            custom_tools: List of tools in OpenAI Chat Completions format
+
+        Returns:
+            List of tools in Gemini-compatible dictionary format
+        """
+        if not custom_tools:
+            return []
+
+        converted_tools = []
+
+        for tool in custom_tools:
+            # Handle OpenAI Chat Completions format with type="function" wrapper
+            if isinstance(tool, dict) and tool.get("type") == "function" and "function" in tool:
+                func_def = tool["function"]
+                converted_tool = {
+                    "name": func_def.get("name", ""),
+                    "description": func_def.get("description", ""),
+                    "parameters": func_def.get("parameters", {}),
+                }
+                converted_tools.append(converted_tool)
+            # Handle already-converted Gemini format (idempotent)
+            elif isinstance(tool, dict) and "name" in tool and "parameters" in tool:
+                # Already in Gemini format, pass through
+                converted_tools.append(tool)
+            else:
+                # Skip unrecognized formats
+                logger.warning(f"[GeminiFormatter] Skipping unrecognized tool format: {tool}")
+
+        return converted_tools
+
+    def _convert_to_function_declarations(self, tools_dicts: List[Dict[str, Any]]) -> List[Any]:
+        """
+        Convert Gemini-format tool dictionaries to FunctionDeclaration objects.
+
+        Args:
+            tools_dicts: List of tool dictionaries in Gemini format
+                [{"name": ..., "description": ..., "parameters": {...}}]
+
+        Returns:
+            List of google.genai.types.FunctionDeclaration objects
+        """
+        if not tools_dicts:
+            return []
+
+        try:
+            from google.genai import types
+        except ImportError:
+            logger.error("[GeminiFormatter] Cannot import google.genai.types for FunctionDeclaration")
+            logger.error("[GeminiFormatter] Falling back to dictionary format")
+            return tools_dicts  # Fallback to dict format
+
+        function_declarations = []
+
+        for tool_dict in tools_dicts:
+            try:
+                # Create Schema object for parameters
+                params = tool_dict.get("parameters", {})
+
+                # Convert parameters to Schema object (recursive)
+                schema = self._build_schema_recursive(params)
+
+                # Create FunctionDeclaration object
+                func_decl = types.FunctionDeclaration(
+                    name=tool_dict.get("name", ""),
+                    description=tool_dict.get("description", ""),
+                    parameters=schema,
+                )
+
+                function_declarations.append(func_decl)
+
+                logger.debug(
+                    f"[GeminiFormatter] Converted tool '{tool_dict.get('name')}' " f"to FunctionDeclaration",
+                )
+
+            except Exception as e:
+                logger.error(
+                    f"[GeminiFormatter] Failed to convert tool to FunctionDeclaration: {e}",
+                )
+                logger.error(f"[GeminiFormatter] Tool dict: {tool_dict}")
+                # Continue processing other tools instead of failing completely
+                continue
+
+        return function_declarations
+
+    def _build_schema_recursive(self, param_schema: Dict[str, Any]) -> Any:
+        """
+        Recursively build a Gemini Schema object from JSON Schema format.
+
+        Handles nested objects, arrays, and all standard JSON Schema types.
+
+        Args:
+            param_schema: JSON Schema dictionary (may be nested)
+
+        Returns:
+            google.genai.types.Schema object
+        """
+        try:
+            from google.genai import types
+        except ImportError:
+            logger.error("[GeminiFormatter] Cannot import google.genai.types")
+            return None
+
+        # Get the type (default to "object" for top-level parameters)
+        param_type = param_schema.get("type", "object")
+        gemini_type = self._convert_json_type_to_gemini_type(param_type)
+
+        # Build base schema kwargs
+        schema_kwargs = {
+            "type": gemini_type,
+        }
+
+        # Add description if present
+        if "description" in param_schema:
+            schema_kwargs["description"] = param_schema["description"]
+
+        # Handle object type with nested properties
+        if param_type == "object" and "properties" in param_schema:
+            schema_kwargs["properties"] = {prop_name: self._build_schema_recursive(prop_schema) for prop_name, prop_schema in param_schema["properties"].items()}
+
+            # Add required fields if present
+            if "required" in param_schema:
+                schema_kwargs["required"] = param_schema["required"]
+
+        # Handle array type with items
+        elif param_type == "array" and "items" in param_schema:
+            schema_kwargs["items"] = self._build_schema_recursive(param_schema["items"])
+
+        # Handle enum if present (for string/number types)
+        if "enum" in param_schema:
+            schema_kwargs["enum"] = param_schema["enum"]
+
+        # Handle format if present (e.g., "date-time", "email", etc.)
+        if "format" in param_schema:
+            schema_kwargs["format"] = param_schema["format"]
+
+        # Handle nullable if present
+        if "nullable" in param_schema:
+            schema_kwargs["nullable"] = param_schema["nullable"]
+
+        return types.Schema(**schema_kwargs)
+
+    def _convert_json_type_to_gemini_type(self, json_type: str) -> Any:
+        """
+        Convert JSON Schema type string to Gemini Type enum.
+
+        Args:
+            json_type: JSON Schema type like "string", "number", "integer", etc.
+
+        Returns:
+            Corresponding google.genai.types.Type enum value
+        """
+        try:
+            from google.genai import types
+        except ImportError:
+            # If we can't import, return string as fallback
+            # This shouldn't happen in practice since _build_schema_recursive checks first
+            return "STRING"
+
+        # Map JSON Schema types to Gemini Type enum
+        type_mapping = {
+            "string": types.Type.STRING,
+            "number": types.Type.NUMBER,
+            "integer": types.Type.INTEGER,
+            "boolean": types.Type.BOOLEAN,
+            "array": types.Type.ARRAY,
+            "object": types.Type.OBJECT,
+        }
+
+        # Return mapped type or default to STRING
+        gemini_type = type_mapping.get(json_type.lower(), types.Type.STRING)
+
+        if json_type.lower() not in type_mapping:
+            logger.warning(
+                f"[GeminiFormatter] Unknown JSON type '{json_type}', defaulting to STRING",
+            )
+
+        return gemini_type

massgen/formatter/_response_formatter.py

@@ -241,6 +241,94 @@ class ResponseFormatter(FormatterBase):
 
         return converted_tools
 
+    def format_custom_tools(self, custom_tools: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """
+        Convert custom tools from RegisteredToolEntry format to Response API format.
+
+        Custom tools are provided as a dictionary where:
+        - Keys are tool names (str)
+        - Values are RegisteredToolEntry objects with:
+          - tool_name: str
+          - schema_def: dict with structure {"type": "function", "function": {...}}
+          - get_extended_schema: property that returns the schema with extensions
+
+        Response API expects: {"type": "function", "name": ..., "description": ..., "parameters": ...}
+
+        Args:
+            custom_tools: Dictionary of tool_name -> RegisteredToolEntry objects
+
+        Returns:
+            List of tools in Response API format
+        """
+        if not custom_tools:
+            return []
+
+        converted_tools = []
+
+        # Handle dictionary format: {tool_name: RegisteredToolEntry, ...}
+        if isinstance(custom_tools, dict):
+            for tool_name, tool_entry in custom_tools.items():
+                # Check if it's a RegisteredToolEntry object with schema_def
+                if hasattr(tool_entry, "schema_def"):
+                    tool_schema = tool_entry.schema_def
+
+                    # Extract function details from Chat Completions format
+                    if tool_schema.get("type") == "function" and "function" in tool_schema:
+                        func = tool_schema["function"]
+                        converted_tools.append(
+                            {
+                                "type": "function",
+                                "name": func.get("name", tool_entry.tool_name if hasattr(tool_entry, "tool_name") else tool_name),
+                                "description": func.get("description", ""),
+                                "parameters": func.get("parameters", {}),
+                            },
+                        )
+                # Check if it has get_extended_schema property
+                elif hasattr(tool_entry, "get_extended_schema"):
+                    tool_schema = tool_entry.get_extended_schema
+
+                    if tool_schema.get("type") == "function" and "function" in tool_schema:
+                        func = tool_schema["function"]
+                        converted_tools.append(
+                            {
+                                "type": "function",
+                                "name": func.get("name", tool_entry.tool_name if hasattr(tool_entry, "tool_name") else tool_name),
+                                "description": func.get("description", ""),
+                                "parameters": func.get("parameters", {}),
+                            },
+                        )
+        # Handle list format for backward compatibility
+        elif isinstance(custom_tools, list):
+            for tool in custom_tools:
+                if hasattr(tool, "schema_def"):
+                    tool_schema = tool.schema_def
+
+                    if tool_schema.get("type") == "function" and "function" in tool_schema:
+                        func = tool_schema["function"]
+                        converted_tools.append(
+                            {
+                                "type": "function",
+                                "name": func.get("name", tool.tool_name),
+                                "description": func.get("description", ""),
+                                "parameters": func.get("parameters", {}),
+                            },
+                        )
+                elif hasattr(tool, "get_extended_schema"):
+                    tool_schema = tool.get_extended_schema
+
+                    if tool_schema.get("type") == "function" and "function" in tool_schema:
+                        func = tool_schema["function"]
+                        converted_tools.append(
+                            {
+                                "type": "function",
+                                "name": func.get("name", tool.tool_name),
+                                "description": func.get("description", ""),
+                                "parameters": func.get("parameters", {}),
+                            },
+                        )
+
+        return converted_tools
+
     def format_mcp_tools(self, mcp_functions: Dict[str, Any]) -> List[Dict[str, Any]]:
         """Convert MCP tools to Response API format (OpenAI function declarations)."""
         if not mcp_functions:

massgen/frontend/coordination_ui.py

@@ -315,7 +315,8 @@ class CoordinationUI:
             #     time.sleep(1.0)
 
             # Get final presentation from winning agent
-            if self.enable_final_presentation and selected_agent and vote_results.get("vote_counts"):
+            # Run final presentation if enabled and there's a selected agent (regardless of votes)
+            if self.enable_final_presentation and selected_agent:
                 # Don't print - let the display handle it
                 # print(f"\n🎤  Final Presentation from {selected_agent}:")
                 # print("=" * 60)
@@ -691,7 +692,8 @@ class CoordinationUI:
             #     time.sleep(1.0)
 
             # Get final presentation from winning agent
-            if self.enable_final_presentation and selected_agent and vote_results.get("vote_counts"):
+            # Run final presentation if enabled and there's a selected agent (regardless of votes)
+            if self.enable_final_presentation and selected_agent:
                 # Don't print - let the display handle it
                 # print(f"\n🎤 Final Presentation from {selected_agent}:")
                 # print("=" * 60)

massgen/logger_config.py

@@ -16,13 +16,21 @@ Color Scheme for Debug Logging:
 """
 
 import inspect
+import subprocess
 import sys
 from datetime import datetime
 from pathlib import Path
 from typing import Any, Optional
 
+import yaml
 from loguru import logger
 
+# Try to import massgen for version info (optional)
+try:
+    import massgen
+except ImportError:
+    massgen = None
+
 # Remove default logger to have full control
 logger.remove()
 
@@ -93,7 +101,12 @@ def get_log_session_dir(turn: Optional[int] = None) -> Path:
     return _LOG_SESSION_DIR
 
 
-def save_execution_metadata(query: str, config_path: Optional[str] = None, config_content: Optional[dict] = None):
+def save_execution_metadata(
+    query: str,
+    config_path: Optional[str] = None,
+    config_content: Optional[dict] = None,
+    cli_args: Optional[dict] = None,
+):
     """Save the query and config metadata to the log directory.
 
     This allows reconstructing what was executed in this session.
@@ -102,9 +115,8 @@ def save_execution_metadata(query: str, config_path: Optional[str] = None, confi
         query: The user's query/prompt
         config_path: Path to the config file that was used (optional)
         config_content: The actual config dictionary (optional)
+        cli_args: Command line arguments as dict (optional)
     """
-    import yaml
-
     log_dir = get_log_session_dir()
 
     # Create a single metadata file with all execution info
@@ -119,6 +131,26 @@ def save_execution_metadata(query: str, config_path: Optional[str] = None, confi
     if config_content:
         metadata["config"] = config_content
 
+    if cli_args:
+        metadata["cli_args"] = cli_args
+
+    # Try to get git information if in a git repository
+    try:
+        git_commit = subprocess.check_output(["git", "rev-parse", "HEAD"], stderr=subprocess.DEVNULL, text=True).strip()
+        git_branch = subprocess.check_output(["git", "rev-parse", "--abbrev-ref", "HEAD"], stderr=subprocess.DEVNULL, text=True).strip()
+        metadata["git"] = {"commit": git_commit, "branch": git_branch}
+    except (subprocess.CalledProcessError, FileNotFoundError):
+        # Not in a git repo or git not available
+        pass
+
+    # Add Python version and package version
+    metadata["python_version"] = sys.version
+    if massgen is not None:
+        metadata["massgen_version"] = getattr(massgen, "__version__", "unknown")
+
+    # Add working directory
+    metadata["working_directory"] = str(Path.cwd())
+
     metadata_file = log_dir / "execution_metadata.yaml"
     try:
         with open(metadata_file, "w", encoding="utf-8") as f: