fast-agent-mcp 0.3.16__py3-none-any.whl → 0.3.17__py3-none-any.whl

fast_agent/agents/mcp_agent.py

@@ -157,7 +157,7 @@ class McpAgent(ABC, ToolAgent):
             if skills_configured:
                 modes.append("skills")
             if shell_flag_requested:
-                modes.append("command switch")
+                modes.append("switch")
             self._shell_access_modes = tuple(modes)
         self._bash_tool = self._shell_runtime.tool
         if self._shell_runtime_enabled:

fast_agent/cli/constants.py

@@ -21,6 +21,8 @@ GO_SPECIFIC_OPTIONS = {
     "-c",
     "--shell",
     "-x",
+    "--skills",
+    "--skills-dir",
 }
 
 # Known subcommands that should not trigger auto-routing

fast_agent/cli/main.py

@@ -8,7 +8,7 @@ from fast_agent.cli.terminal import Application
 from fast_agent.ui.console import console as shared_console
 
 app = typer.Typer(
-    help="fast-agent - Build effective agents using Model Context Protocol",
+    help="Use `fast-agent go --help` for interactive shell arguments and options.",
     add_completion=False,  # We'll add this later when we have more commands
 )
 

fast_agent/interfaces.py

@@ -18,6 +18,7 @@ from typing import (
     Type,
     TypeVar,
     Union,
+    runtime_checkable,
 )
 
 from a2a.types import AgentCard
@@ -59,6 +60,7 @@ class ModelFactoryFunctionProtocol(Protocol):
     def __call__(self, model: str | None = None) -> LLMFactoryProtocol: ...
 
 
+@runtime_checkable
 class FastAgentLLMProtocol(Protocol):
     """Protocol defining the interface for LLMs"""
 
@@ -111,6 +113,7 @@ class FastAgentLLMProtocol(Protocol):
     def clear(self, *, clear_prompts: bool = False) -> None: ...
 
 
+@runtime_checkable
 class LlmAgentProtocol(Protocol):
     """Protocol defining the minimal interface for LLM agents."""
 
@@ -132,6 +135,7 @@ class LlmAgentProtocol(Protocol):
     def pop_last_message(self) -> PromptMessageExtended | None: ...
 
 
+@runtime_checkable
 class AgentProtocol(LlmAgentProtocol, Protocol):
     """Standard agent interface with flexible input types."""
 

fast_agent/llm/model_database.py

@@ -87,7 +87,7 @@ class ModelDatabase:
     )
 
     GEMINI_PRO = ModelParameters(
-        context_window=2097152, max_output_tokens=8192, tokenizes=GOOGLE_MULTIMODAL
+        context_window=1_048_576, max_output_tokens=65_536, tokenizes=GOOGLE_MULTIMODAL
     )
 
     QWEN_STANDARD = ModelParameters(
@@ -245,6 +245,9 @@ class ModelDatabase:
         "gemini-2.5-pro-preview": GEMINI_2_5_PRO,
         "gemini-2.5-flash-preview-05-20": GEMINI_FLASH,
         "gemini-2.5-pro-preview-05-06": GEMINI_PRO,
+        "gemini-2.5-pro": GEMINI_PRO,
+        "gemini-2.5-flash-preview-09-2025": GEMINI_FLASH,
+        "gemini-2.5-flash": GEMINI_FLASH,
         # xAI Grok Models
         "grok-4-fast-reasoning": GROK_4_VLM,
         "grok-4-fast-non-reasoning": GROK_4_VLM,

fast_agent/llm/model_factory.py

@@ -90,7 +90,9 @@ class ModelFactory:
         "deepseek-chat": Provider.DEEPSEEK,
         "gemini-2.0-flash": Provider.GOOGLE,
         "gemini-2.5-flash-preview-05-20": Provider.GOOGLE,
+        "gemini-2.5-flash-preview-09-2025": Provider.GOOGLE,
         "gemini-2.5-pro-preview-05-06": Provider.GOOGLE,
+        "gemini-2.5-pro": Provider.GOOGLE,
         "grok-4": Provider.XAI,
         "grok-4-0709": Provider.XAI,
         "grok-3": Provider.XAI,
@@ -120,8 +122,8 @@ class ModelFactory:
         "deepseekv3": "deepseek-chat",
         "deepseek": "deepseek-chat",
         "gemini2": "gemini-2.0-flash",
-        "gemini25": "gemini-2.5-flash-preview-05-20",
-        "gemini25pro": "gemini-2.5-pro-preview-05-06",
+        "gemini25": "gemini-2.5-flash-preview-09-2025",
+        "gemini25pro": "gemini-2.5-pro",
         "kimi": "groq.moonshotai/kimi-k2-instruct-0905",
         "gpt-oss": "groq.openai/gpt-oss-120b",
         "gpt-oss-20b": "groq.openai/gpt-oss-20b",

fast_agent/llm/model_info.py

@@ -8,14 +8,15 @@ capabilities (Text/Document/Vision), backed by the model database.
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, List, Optional, Union
+from typing import TYPE_CHECKING, List, Optional
 
 from fast_agent.llm.model_database import ModelDatabase
+from fast_agent.llm.model_factory import ModelFactory
 from fast_agent.llm.provider_types import Provider
 
 if TYPE_CHECKING:
     # Import behind TYPE_CHECKING to avoid import cycles at runtime
-    from fast_agent.interfaces import AgentProtocol, FastAgentLLMProtocol
+    from fast_agent.interfaces import FastAgentLLMProtocol
 
 
 @dataclass(frozen=True)
@@ -32,16 +33,24 @@ class ModelInfo:
 
     @property
     def supports_text(self) -> bool:
+        if "text/plain" in (self.tokenizes or []):
+            return True
         return ModelDatabase.supports_mime(self.name, "text/plain")
 
     @property
     def supports_document(self) -> bool:
         # Document support currently keyed off PDF support
+        if "application/pdf" in (self.tokenizes or []):
+            return True
         return ModelDatabase.supports_mime(self.name, "pdf")
 
     @property
     def supports_vision(self) -> bool:
         # Any common image format indicates vision support
+        tokenizes = self.tokenizes or []
+        if any(mt in tokenizes for mt in ("image/jpeg", "image/png", "image/webp")):
+            return True
+
         return any(
             ModelDatabase.supports_mime(self.name, mt)
             for mt in ("image/jpeg", "image/png", "image/webp")
@@ -62,14 +71,15 @@ class ModelInfo:
 
     @classmethod
     def from_name(cls, name: str, provider: Provider | None = None) -> Optional["ModelInfo"]:
-        params = ModelDatabase.get_model_params(name)
+        canonical_name = ModelFactory.MODEL_ALIASES.get(name, name)
+        params = ModelDatabase.get_model_params(canonical_name)
         if not params:
             # Unknown model: return a conservative default that supports text only.
             # This matches the desired behavior for TDV display fallbacks.
             if provider is None:
                 provider = Provider.GENERIC
             return ModelInfo(
-                name=name,
+                name=canonical_name,
                 provider=provider,
                 context_window=None,
                 max_output_tokens=None,
@@ -78,49 +88,15 @@ class ModelInfo:
                 reasoning=None,
             )
 
+        if provider is None:
+            provider = ModelFactory.DEFAULT_PROVIDERS.get(canonical_name, Provider.GENERIC)
+
         return ModelInfo(
-            name=name,
-            provider=provider or Provider.GENERIC,
+            name=canonical_name,
+            provider=provider,
             context_window=params.context_window,
             max_output_tokens=params.max_output_tokens,
             tokenizes=params.tokenizes,
             json_mode=params.json_mode,
             reasoning=params.reasoning,
         )
-
-
-def get_model_info(
-    subject: Union["AgentProtocol", "FastAgentLLMProtocol", str, None],
-    provider: Provider | None = None,
-) -> Optional[ModelInfo]:
-    """Resolve a ModelInfo from an Agent, LLM, or model name.
-
-    Keeps the public API small while enabling type-safe access to model
-    capabilities across the codebase.
-    """
-    if subject is None:
-        return None
-
-    # Agent → LLM
-    try:
-        from fast_agent.interfaces import AgentProtocol as _AgentProtocol
-    except Exception:
-        _AgentProtocol = None  # type: ignore
-
-    if _AgentProtocol and isinstance(subject, _AgentProtocol):  # type: ignore[arg-type]
-        return ModelInfo.from_llm(subject.llm)
-
-    # LLM → ModelInfo
-    try:
-        from fast_agent.interfaces import FastAgentLLMProtocol as _LLMProtocol
-    except Exception:
-        _LLMProtocol = None  # type: ignore
-
-    if _LLMProtocol and isinstance(subject, _LLMProtocol):  # type: ignore[arg-type]
-        return ModelInfo.from_llm(subject)
-
-    # String model name
-    if isinstance(subject, str):
-        return ModelInfo.from_name(subject, provider)
-
-    return None

fast_agent/llm/provider/google/llm_google_native.py

@@ -1,3 +1,4 @@
+import json
 import secrets
 from typing import Dict, List
 
@@ -49,8 +50,6 @@ class GoogleNativeLLM(FastAgentLLM[types.Content, types.Content]):
 
     def __init__(self, *args, **kwargs) -> None:
         super().__init__(*args, provider=Provider.GOOGLE, **kwargs)
-        # Initialize the google.genai client
-        self._google_client = self._initialize_google_client()
         # Initialize the converter
         self._converter = GoogleConverter()
 
@@ -109,6 +108,218 @@ class GoogleNativeLLM(FastAgentLLM[types.Content, types.Content]):
             # Include other relevant default parameters
         )
 
+    async def _stream_generate_content(
+        self,
+        *,
+        model: str,
+        contents: List[types.Content],
+        config: types.GenerateContentConfig,
+        client: genai.Client,
+    ) -> types.GenerateContentResponse | None:
+        """Stream Gemini responses and return the final aggregated completion."""
+        try:
+            response_stream = await client.aio.models.generate_content_stream(
+                model=model,
+                contents=contents,
+                config=config,
+            )
+        except AttributeError:
+            # Older SDKs might not expose streaming; fall back to non-streaming.
+            return None
+        except errors.APIError:
+            raise
+        except Exception as exc:  # pragma: no cover - defensive fallback
+            self.logger.warning(
+                "Google streaming failed during setup; falling back to non-streaming",
+                exc_info=exc,
+            )
+            return None
+
+        return await self._consume_google_stream(response_stream, model=model)
+
+    async def _consume_google_stream(
+        self,
+        response_stream,
+        *,
+        model: str,
+    ) -> types.GenerateContentResponse | None:
+        """Consume the async streaming iterator and aggregate the final response."""
+        estimated_tokens = 0
+        timeline: List[tuple[str, int | None, str]] = []
+        tool_streams: Dict[int, Dict[str, str]] = {}
+        active_tool_index: int | None = None
+        tool_counter = 0
+        usage_metadata = None
+        last_chunk: types.GenerateContentResponse | None = None
+
+        try:
+            async for chunk in response_stream:
+                last_chunk = chunk
+                if getattr(chunk, "usage_metadata", None):
+                    usage_metadata = chunk.usage_metadata
+
+                if not getattr(chunk, "candidates", None):
+                    continue
+
+                candidate = chunk.candidates[0]
+                content = getattr(candidate, "content", None)
+                if content is None or not getattr(content, "parts", None):
+                    continue
+
+                for part in content.parts:
+                    if getattr(part, "text", None):
+                        text = part.text or ""
+                        if text:
+                            if timeline and timeline[-1][0] == "text":
+                                prev_type, prev_index, prev_text = timeline[-1]
+                                timeline[-1] = (prev_type, prev_index, prev_text + text)
+                            else:
+                                timeline.append(("text", None, text))
+                            estimated_tokens = self._update_streaming_progress(
+                                text,
+                                model,
+                                estimated_tokens,
+                            )
+                            self._notify_tool_stream_listeners(
+                                "text",
+                                {
+                                    "chunk": text,
+                                    "streams_arguments": False,
+                                },
+                            )
+
+                    if getattr(part, "function_call", None):
+                        function_call = part.function_call
+                        name = getattr(function_call, "name", None) or "tool"
+                        args = getattr(function_call, "args", None) or {}
+
+                        if active_tool_index is None:
+                            active_tool_index = tool_counter
+                            tool_counter += 1
+                            tool_use_id = f"tool_{self.chat_turn()}_{active_tool_index}"
+                            tool_streams[active_tool_index] = {
+                                "name": name,
+                                "tool_use_id": tool_use_id,
+                                "buffer": "",
+                            }
+                            self._notify_tool_stream_listeners(
+                                "start",
+                                {
+                                    "tool_name": name,
+                                    "tool_use_id": tool_use_id,
+                                    "index": active_tool_index,
+                                    "streams_arguments": False,
+                                },
+                            )
+                            timeline.append(("tool_call", active_tool_index, ""))
+
+                        stream_info = tool_streams.get(active_tool_index)
+                        if not stream_info:
+                            continue
+
+                        try:
+                            serialized_args = json.dumps(args, separators=(",", ":"))
+                        except Exception:
+                            serialized_args = str(args)
+
+                        previous = stream_info.get("buffer", "")
+                        if isinstance(previous, str) and serialized_args.startswith(previous):
+                            delta = serialized_args[len(previous) :]
+                        else:
+                            delta = serialized_args
+                        stream_info["buffer"] = serialized_args
+
+                        if delta:
+                            self._notify_tool_stream_listeners(
+                                "delta",
+                                {
+                                    "tool_name": stream_info["name"],
+                                    "tool_use_id": stream_info["tool_use_id"],
+                                    "index": active_tool_index,
+                                    "chunk": delta,
+                                    "streams_arguments": False,
+                                },
+                            )
+
+                finish_reason = getattr(candidate, "finish_reason", None)
+                if finish_reason:
+                    finish_value = str(finish_reason).split(".")[-1].upper()
+                    if finish_value in {"FUNCTION_CALL", "STOP"} and active_tool_index is not None:
+                        stream_info = tool_streams.get(active_tool_index)
+                        if stream_info:
+                            self._notify_tool_stream_listeners(
+                                "stop",
+                                {
+                                    "tool_name": stream_info["name"],
+                                    "tool_use_id": stream_info["tool_use_id"],
+                                    "index": active_tool_index,
+                                    "streams_arguments": False,
+                                },
+                            )
+                        active_tool_index = None
+        finally:
+            stream_close = getattr(response_stream, "aclose", None)
+            if callable(stream_close):
+                try:
+                    await stream_close()
+                except Exception:
+                    pass
+
+        if active_tool_index is not None:
+            stream_info = tool_streams.get(active_tool_index)
+            if stream_info:
+                self._notify_tool_stream_listeners(
+                    "stop",
+                    {
+                        "tool_name": stream_info["name"],
+                        "tool_use_id": stream_info["tool_use_id"],
+                        "index": active_tool_index,
+                        "streams_arguments": False,
+                    },
+                )
+
+        if not timeline and last_chunk is None:
+            return None
+
+        final_parts: List[types.Part] = []
+        for entry_type, index, payload in timeline:
+            if entry_type == "text":
+                final_parts.append(types.Part.from_text(text=payload))
+            elif entry_type == "tool_call" and index is not None:
+                stream_info = tool_streams.get(index)
+                if not stream_info:
+                    continue
+                buffer = stream_info.get("buffer", "")
+                try:
+                    args_obj = json.loads(buffer) if buffer else {}
+                except json.JSONDecodeError:
+                    args_obj = {"__raw": buffer}
+                final_parts.append(
+                    types.Part.from_function_call(
+                        name=str(stream_info.get("name") or "tool"),
+                        args=args_obj,
+                    )
+                )
+
+        final_content = types.Content(role="model", parts=final_parts)
+
+        if last_chunk is not None:
+            final_response = last_chunk.model_copy(deep=True)
+            if getattr(final_response, "candidates", None):
+                final_candidate = final_response.candidates[0]
+                final_candidate.content = final_content
+            else:
+                final_response.candidates = [types.Candidate(content=final_content)]
+        else:
+            final_response = types.GenerateContentResponse(
+                candidates=[types.Candidate(content=final_content)]
+            )
+
+        if usage_metadata:
+            final_response.usage_metadata = usage_metadata
+
+        return final_response
+
     async def _google_completion(
         self,
         message: List[types.Content] | None,
@@ -163,13 +374,24 @@ class GoogleNativeLLM(FastAgentLLM[types.Content, types.Content]):
             )
 
         # 3. Call the google.genai API
+        client = self._initialize_google_client()
         try:
             # Use the async client
-            api_response = await self._google_client.aio.models.generate_content(
-                model=request_params.model,
-                contents=conversation_history,  # Full conversational context for this turn
-                config=generate_content_config,
-            )
+            api_response = None
+            streaming_supported = response_schema is None and response_mime_type is None
+            if streaming_supported:
+                api_response = await self._stream_generate_content(
+                    model=request_params.model,
+                    contents=conversation_history,
+                    config=generate_content_config,
+                    client=client,
+                )
+            if api_response is None:
+                api_response = await client.aio.models.generate_content(
+                    model=request_params.model,
+                    contents=conversation_history,  # Full conversational context for this turn
+                    config=generate_content_config,
+                )
             self.logger.debug("Google generate_content response:", data=api_response)
 
             # Track usage if response is valid and has usage data
@@ -195,6 +417,15 @@ class GoogleNativeLLM(FastAgentLLM[types.Content, types.Content]):
             self.logger.error(f"Error during Google generate_content call: {e}")
             # Decide how to handle other exceptions - potentially re-raise or return an error message
             raise e
+        finally:
+            try:
+                await client.aio.aclose()
+            except Exception:
+                pass
+            try:
+                client.close()
+            except Exception:
+                pass
 
         # 4. Process the API response
         if not api_response.candidates: