code-puppy 0.0.302__py3-none-any.whl → 0.0.335__py3-none-any.whl

code_puppy/agents/base_agent.py

@@ -7,7 +7,19 @@ import signal
 import threading
 import uuid
 from abc import ABC, abstractmethod
-from typing import Any, Callable, Dict, List, Optional, Sequence, Set, Tuple, Union
+from collections.abc import AsyncIterable
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    List,
+    Optional,
+    Sequence,
+    Set,
+    Tuple,
+    Type,
+    Union,
+)
 
 import mcp
 import pydantic
@@ -18,6 +30,7 @@ from pydantic_ai import (
     BinaryContent,
     DocumentUrl,
     ImageUrl,
+    PartEndEvent,
     RunContext,
     UsageLimitExceeded,
     UsageLimits,
@@ -33,6 +46,7 @@ from pydantic_ai.messages import (
     ToolReturn,
     ToolReturnPart,
 )
+from rich.text import Text
 
 # Consolidated relative imports
 from code_puppy.config import (
@@ -44,11 +58,10 @@ from code_puppy.config import (
     get_protected_token_count,
     get_use_dbos,
     get_value,
-    load_mcp_server_configs,
 )
 from code_puppy.error_logging import log_error
 from code_puppy.keymap import cancel_agent_uses_signal, get_cancel_agent_char_code
-from code_puppy.mcp_ import ServerConfig, get_mcp_manager
+from code_puppy.mcp_ import get_mcp_manager
 from code_puppy.messaging import (
     emit_error,
     emit_info,
@@ -87,6 +100,9 @@ class BaseAgent(ABC):
         # Cache for MCP tool definitions (for token estimation)
         # This is populated after the first successful run when MCP tools are retrieved
         self._mcp_tool_definitions_cache: List[Dict[str, Any]] = []
+        # Shared console for streaming output - should be set by cli_runner
+        # to avoid conflicts between spinner's Live display and response streaming
+        self._console: Optional[Any] = None
 
     @property
     @abstractmethod
@@ -364,7 +380,9 @@ class BaseAgent(ABC):
         # fixed instructions. For other models, count the full system prompt.
         try:
             from code_puppy.model_utils import (
+                get_chatgpt_codex_instructions,
                 get_claude_code_instructions,
+                is_chatgpt_codex_model,
                 is_claude_code_model,
             )
 
@@ -376,6 +394,11 @@ class BaseAgent(ABC):
                 # The full system prompt is already in the message history
                 instructions = get_claude_code_instructions()
                 total_tokens += self.estimate_token_count(instructions)
+            elif is_chatgpt_codex_model(model_name):
+                # For ChatGPT Codex models, only count the short fixed instructions
+                # The full system prompt is already in the message history
+                instructions = get_chatgpt_codex_instructions()
+                total_tokens += self.estimate_token_count(instructions)
             else:
                 # For other models, count the full system prompt
                 system_prompt = self.get_system_prompt()
@@ -979,45 +1002,31 @@ class BaseAgent(ABC):
         return self._puppy_rules
 
     def load_mcp_servers(self, extra_headers: Optional[Dict[str, str]] = None):
-        """Load MCP servers through the manager and return pydantic-ai compatible servers."""
+        """Load MCP servers through the manager and return pydantic-ai compatible servers.
+
+        Note: The manager automatically syncs from mcp_servers.json during initialization,
+        so we don't need to sync here. Use reload_mcp_servers() to force a re-sync.
+        """
 
         mcp_disabled = get_value("disable_mcp_servers")
         if mcp_disabled and str(mcp_disabled).lower() in ("1", "true", "yes", "on"):
             return []
 
         manager = get_mcp_manager()
-        configs = load_mcp_server_configs()
-        if not configs:
-            existing_servers = manager.list_servers()
-            if not existing_servers:
-                return []
-        else:
-            for name, conf in configs.items():
-                try:
-                    server_config = ServerConfig(
-                        id=conf.get("id", f"{name}_{hash(name)}"),
-                        name=name,
-                        type=conf.get("type", "sse"),
-                        enabled=conf.get("enabled", True),
-                        config=conf,
-                    )
-                    existing = manager.get_server_by_name(name)
-                    if not existing:
-                        manager.register_server(server_config)
-                    else:
-                        if existing.config != server_config.config:
-                            manager.update_server(existing.id, server_config)
-                except Exception:
-                    continue
-
         return manager.get_servers_for_agent()
 
     def reload_mcp_servers(self):
-        """Reload MCP servers and return updated servers."""
+        """Reload MCP servers and return updated servers.
+
+        Forces a re-sync from mcp_servers.json to pick up any configuration changes.
+        """
         # Clear the MCP tool cache when servers are reloaded
         self._mcp_tool_definitions_cache = []
-        self.load_mcp_servers()
+
+        # Force re-sync from mcp_servers.json
         manager = get_mcp_manager()
+        manager.sync_from_config()
+
         return manager.get_servers_for_agent()
 
     def _load_model_with_fallback(
@@ -1173,7 +1182,9 @@ class BaseAgent(ABC):
 
         if len(filtered_mcp_servers) != len(mcp_servers):
             emit_info(
-                f"[dim]Filtered {len(mcp_servers) - len(filtered_mcp_servers)} conflicting MCP tools[/dim]"
+                Text.from_markup(
+                    f"[dim]Filtered {len(mcp_servers) - len(filtered_mcp_servers)} conflicting MCP tools[/dim]"
+                )
             )
 
         self._last_model_name = resolved_model_name
@@ -1230,6 +1241,74 @@ class BaseAgent(ABC):
             self._mcp_servers = mcp_servers
         return self._code_generation_agent
 
+    def _create_agent_with_output_type(self, output_type: Type[Any]) -> PydanticAgent:
+        """Create a temporary agent configured with a custom output_type.
+
+        This is used when structured output is requested via run_with_mcp.
+        The agent is created fresh with the same configuration as the main agent
+        but with the specified output_type instead of str.
+
+        Args:
+            output_type: The Pydantic model or type for structured output.
+
+        Returns:
+            A configured PydanticAgent (or DBOSAgent wrapper) with the custom output_type.
+        """
+        from code_puppy.model_utils import prepare_prompt_for_model
+        from code_puppy.tools import register_tools_for_agent
+
+        model_name = self.get_model_name()
+        models_config = ModelFactory.load_config()
+        model, resolved_model_name = self._load_model_with_fallback(
+            model_name, models_config, str(uuid.uuid4())
+        )
+
+        instructions = self.get_system_prompt()
+        puppy_rules = self.load_puppy_rules()
+        if puppy_rules:
+            instructions += f"\n{puppy_rules}"
+
+        mcp_servers = getattr(self, "_mcp_servers", []) or []
+        model_settings = make_model_settings(resolved_model_name)
+
+        prepared = prepare_prompt_for_model(
+            model_name, instructions, "", prepend_system_to_user=False
+        )
+        instructions = prepared.instructions
+
+        global _reload_count
+        _reload_count += 1
+
+        if get_use_dbos():
+            temp_agent = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=output_type,
+                retries=3,
+                toolsets=[],
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(temp_agent, agent_tools)
+            dbos_agent = DBOSAgent(
+                temp_agent, name=f"{self.name}-structured-{_reload_count}"
+            )
+            return dbos_agent
+        else:
+            temp_agent = PydanticAgent(
+                model=model,
+                instructions=instructions,
+                output_type=output_type,
+                retries=3,
+                toolsets=mcp_servers,
+                history_processors=[self.message_history_accumulator],
+                model_settings=model_settings,
+            )
+            agent_tools = self.get_available_tools()
+            register_tools_for_agent(temp_agent, agent_tools)
+            return temp_agent
+
     # It's okay to decorate it with DBOS.step even if not using DBOS; the decorator is a no-op in that case.
     @DBOS.step()
     def message_history_accumulator(self, ctx: RunContext, messages: List[Any]):
@@ -1255,6 +1334,216 @@ class BaseAgent(ABC):
             self.set_message_history(result_messages_filtered_empty_thinking)
         return self.get_message_history()
 
+    async def _event_stream_handler(
+        self, ctx: RunContext, events: AsyncIterable[Any]
+    ) -> None:
+        """Handle streaming events from the agent run.
+
+        This method processes streaming events and emits TextPart, ThinkingPart,
+        and ToolCallPart content with styled banners/tokens as they stream in.
+
+        Args:
+            ctx: The run context.
+            events: Async iterable of streaming events (PartStartEvent, PartDeltaEvent, etc.).
+        """
+        from pydantic_ai import PartDeltaEvent, PartStartEvent
+        from pydantic_ai.messages import (
+            TextPartDelta,
+            ThinkingPartDelta,
+            ToolCallPartDelta,
+        )
+        from rich.console import Console
+        from rich.markdown import Markdown
+        from rich.markup import escape
+
+        from code_puppy.messaging.spinner import pause_all_spinners
+
+        # IMPORTANT: Use the shared console (set by cli_runner) to avoid conflicts
+        # with the spinner's Live display. Multiple Console instances with separate
+        # Live displays cause cursor positioning chaos and line duplication.
+        if self._console is not None:
+            console = self._console
+        else:
+            # Fallback if console not set (shouldn't happen in normal use)
+            console = Console()
+
+        # Track which part indices we're currently streaming (for Text/Thinking/Tool parts)
+        streaming_parts: set[int] = set()
+        thinking_parts: set[int] = (
+            set()
+        )  # Track which parts are thinking (for dim style)
+        text_parts: set[int] = set()  # Track which parts are text
+        tool_parts: set[int] = set()  # Track which parts are tool calls
+        banner_printed: set[int] = set()  # Track if banner was already printed
+        text_buffer: dict[int, list[str]] = {}  # Buffer text for final markdown render
+        token_count: dict[int, int] = {}  # Track token count per text/tool part
+        did_stream_anything = False  # Track if we streamed any content
+
+        def _print_thinking_banner() -> None:
+            """Print the THINKING banner with spinner pause and line clear."""
+            nonlocal did_stream_anything
+            import time
+
+            from code_puppy.config import get_banner_color
+
+            pause_all_spinners()
+            time.sleep(0.1)  # Delay to let spinner fully clear
+            # Clear line and print newline before banner
+            console.print(" " * 50, end="\r")
+            console.print()  # Newline before banner
+            # Bold banner with configurable color and lightning bolt
+            thinking_color = get_banner_color("thinking")
+            console.print(
+                Text.from_markup(
+                    f"[bold white on {thinking_color}] THINKING [/bold white on {thinking_color}] [dim]⚡ "
+                ),
+                end="",
+            )
+            did_stream_anything = True
+
+        def _print_response_banner() -> None:
+            """Print the AGENT RESPONSE banner with spinner pause and line clear."""
+            nonlocal did_stream_anything
+            import time
+
+            from code_puppy.config import get_banner_color
+
+            pause_all_spinners()
+            time.sleep(0.1)  # Delay to let spinner fully clear
+            # Clear line and print newline before banner
+            console.print(" " * 50, end="\r")
+            console.print()  # Newline before banner
+            response_color = get_banner_color("agent_response")
+            console.print(
+                Text.from_markup(
+                    f"[bold white on {response_color}] AGENT RESPONSE [/bold white on {response_color}]"
+                )
+            )
+            did_stream_anything = True
+
+        async for event in events:
+            # PartStartEvent - register the part but defer banner until content arrives
+            if isinstance(event, PartStartEvent):
+                part = event.part
+                if isinstance(part, ThinkingPart):
+                    streaming_parts.add(event.index)
+                    thinking_parts.add(event.index)
+                    # If there's initial content, print banner + content now
+                    if part.content and part.content.strip():
+                        _print_thinking_banner()
+                        escaped = escape(part.content)
+                        console.print(f"[dim]{escaped}[/dim]", end="")
+                        banner_printed.add(event.index)
+                elif isinstance(part, TextPart):
+                    streaming_parts.add(event.index)
+                    text_parts.add(event.index)
+                    text_buffer[event.index] = []  # Initialize buffer
+                    token_count[event.index] = 0  # Initialize token counter
+                    # Buffer initial content if present
+                    if part.content and part.content.strip():
+                        text_buffer[event.index].append(part.content)
+                        # Count chunks (each part counts as 1)
+                        token_count[event.index] += 1
+                elif isinstance(part, ToolCallPart):
+                    streaming_parts.add(event.index)
+                    tool_parts.add(event.index)
+                    token_count[event.index] = 0  # Initialize token counter
+                    # Track tool name for display
+                    banner_printed.add(
+                        event.index
+                    )  # Use banner_printed to track if we've shown tool info
+
+            # PartDeltaEvent - stream the content as it arrives
+            elif isinstance(event, PartDeltaEvent):
+                if event.index in streaming_parts:
+                    delta = event.delta
+                    if isinstance(delta, (TextPartDelta, ThinkingPartDelta)):
+                        if delta.content_delta:
+                            # For text parts, show token counter then render at end
+                            if event.index in text_parts:
+                                # Print banner on first content
+                                if event.index not in banner_printed:
+                                    _print_response_banner()
+                                    banner_printed.add(event.index)
+                                # Accumulate text for final markdown render
+                                text_buffer[event.index].append(delta.content_delta)
+                                # Count chunks received
+                                token_count[event.index] += 1
+                                # Update chunk counter in place (single line)
+                                count = token_count[event.index]
+                                console.print(
+                                    f"  ⏳ Receiving... {count} chunks   ",
+                                    end="\r",
+                                )
+                            else:
+                                # For thinking parts, stream immediately (dim)
+                                if event.index not in banner_printed:
+                                    _print_thinking_banner()
+                                    banner_printed.add(event.index)
+                                escaped = escape(delta.content_delta)
+                                console.print(f"[dim]{escaped}[/dim]", end="")
+                    elif isinstance(delta, ToolCallPartDelta):
+                        # For tool calls, count chunks received
+                        token_count[event.index] += 1
+                        # Get tool name if available
+                        tool_name = getattr(delta, "tool_name_delta", "")
+                        count = token_count[event.index]
+                        # Display with tool wrench icon and tool name
+                        if tool_name:
+                            console.print(
+                                f"  🔧 Calling {tool_name}... {count} chunks   ",
+                                end="\r",
+                            )
+                        else:
+                            console.print(
+                                f"  🔧 Calling tool... {count} chunks   ",
+                                end="\r",
+                            )
+
+            # PartEndEvent - finish the streaming with a newline
+            elif isinstance(event, PartEndEvent):
+                if event.index in streaming_parts:
+                    # For text parts, clear counter line and render markdown
+                    if event.index in text_parts:
+                        # Clear the chunk counter line by printing spaces and returning
+                        console.print(" " * 50, end="\r")
+                        # Render the final markdown nicely
+                        if event.index in text_buffer:
+                            try:
+                                final_content = "".join(text_buffer[event.index])
+                                if final_content.strip():
+                                    console.print(Markdown(final_content))
+                            except Exception:
+                                pass
+                            del text_buffer[event.index]
+                    # For tool parts, clear the chunk counter line
+                    elif event.index in tool_parts:
+                        # Clear the chunk counter line by printing spaces and returning
+                        console.print(" " * 50, end="\r")
+                    # For thinking parts, just print newline
+                    elif event.index in banner_printed:
+                        console.print()  # Final newline after streaming
+
+                    # Clean up token count
+                    token_count.pop(event.index, None)
+                    # Clean up all tracking sets
+                    streaming_parts.discard(event.index)
+                    thinking_parts.discard(event.index)
+                    text_parts.discard(event.index)
+                    tool_parts.discard(event.index)
+                    banner_printed.discard(event.index)
+
+                    # Resume spinner if next part is NOT text/thinking/tool (avoid race condition)
+                    # If next part is None or handled differently, it's safe to resume
+                    # Note: spinner itself handles blank line before appearing
+                    from code_puppy.messaging.spinner import resume_all_spinners
+
+                    next_kind = getattr(event, "next_part_kind", None)
+                    if next_kind not in ("text", "thinking", "tool-call"):
+                        resume_all_spinners()
+
+        # Spinner is resumed in PartEndEvent when appropriate (based on next_part_kind)
+
     def _spawn_ctrl_x_key_listener(
         self,
         stop_event: threading.Event,
@@ -1409,6 +1698,7 @@ class BaseAgent(ABC):
         *,
         attachments: Optional[Sequence[BinaryContent]] = None,
         link_attachments: Optional[Sequence[Union[ImageUrl, DocumentUrl]]] = None,
+        output_type: Optional[Type[Any]] = None,
         **kwargs,
     ) -> Any:
         """Run the agent with MCP servers, attachments, and full cancellation support.
@@ -1417,10 +1707,13 @@ class BaseAgent(ABC):
             prompt: Primary user prompt text (may be empty when attachments present).
             attachments: Local binary payloads (e.g., dragged images) to include.
             link_attachments: Remote assets (image/document URLs) to include.
+            output_type: Optional Pydantic model or type for structured output.
+                When provided, creates a temporary agent configured to return
+                this type instead of the default string output.
             **kwargs: Additional arguments forwarded to `pydantic_ai.Agent.run`.
 
         Returns:
-            The agent's response.
+            The agent's response (typed according to output_type if specified).
 
         Raises:
             asyncio.CancelledError: When execution is cancelled by user.
@@ -1444,10 +1737,17 @@ class BaseAgent(ABC):
         pydantic_agent = (
             self._code_generation_agent or self.reload_code_generation_agent()
         )
-        # Handle claude-code models: prepend system prompt to first user message
-        from code_puppy.model_utils import is_claude_code_model
 
-        if is_claude_code_model(self.get_model_name()):
+        # If a custom output_type is specified, create a temporary agent with that type
+        if output_type is not None:
+            pydantic_agent = self._create_agent_with_output_type(output_type)
+
+        # Handle claude-code and chatgpt-codex models: prepend system prompt to first user message
+        from code_puppy.model_utils import is_chatgpt_codex_model, is_claude_code_model
+
+        if is_claude_code_model(self.get_model_name()) or is_chatgpt_codex_model(
+            self.get_model_name()
+        ):
             if len(self.get_message_history()) == 0:
                 system_prompt = self.get_system_prompt()
                 puppy_rules = self.load_puppy_rules()
@@ -1511,6 +1811,7 @@ class BaseAgent(ABC):
                                 prompt_payload,
                                 message_history=self.get_message_history(),
                                 usage_limits=usage_limits,
+                                event_stream_handler=self._event_stream_handler,
                                 **kwargs,
                             )
                     finally:
@@ -1523,6 +1824,7 @@ class BaseAgent(ABC):
                             prompt_payload,
                             message_history=self.get_message_history(),
                             usage_limits=usage_limits,
+                            event_stream_handler=self._event_stream_handler,
                             **kwargs,
                         )
                 else:
@@ -1531,6 +1833,7 @@ class BaseAgent(ABC):
                         prompt_payload,
                         message_history=self.get_message_history(),
                         usage_limits=usage_limits,
+                        event_stream_handler=self._event_stream_handler,
                         **kwargs,
                     )
                 return result_
@@ -1635,7 +1938,12 @@ class BaseAgent(ABC):
         def graceful_sigint_handler(_sig, _frame):
             # When using keyboard-based cancel, SIGINT should be a no-op
             # (just show a hint to user about the configured cancel key)
+            # Also reset terminal to prevent bricking on Windows+uvx
             from code_puppy.keymap import get_cancel_agent_display_name
+            from code_puppy.terminal_utils import reset_windows_terminal_full
+
+            # Reset terminal state first to prevent bricking
+            reset_windows_terminal_full()
 
             cancel_key = get_cancel_agent_display_name()
             emit_info(f"Use {cancel_key} to cancel the agent task.")