batch parallelization, media attach (#793)

* batch parallelization, media attach

* deps

* smart prompt

Author: evalstate

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "fast-agent-mcp"
-version = "0.7.3"
+version = "0.7.4"
 description = "Define, Prompt and Test MCP enabled Agents and Workflows"
 readme = "README.md"
 license = { file = "LICENSE" }
@@ -15,14 +15,14 @@ classifiers = [
 requires-python = ">=3.13.5,<3.15"
 dependencies = [
     "fastapi==0.136.1",
-    "fastmcp==3.2.4",
+    "fastmcp==3.3.1",
     "mcp==1.27.1",
-    "pydantic-settings==2.13.0",
-    "pydantic==2.13.3",
+    "pydantic-settings==2.14.1",
+    "pydantic==2.13.4",
     "pyyaml==6.0.3",
     "rich==15.0.0",
     "typer==0.25.1",
-    "anthropic[vertex]==0.100.0",
+    "anthropic[vertex]==0.102.0",
     "openai[aiohttp]==2.36.0",
     "prompt-toolkit==3.0.52",
     "aiohttp==3.13.5",
@@ -32,21 +32,21 @@ dependencies = [
     "opentelemetry-instrumentation-anthropic==0.52.1; python_version >= '3.10' and python_version < '4.0'",
     "opentelemetry-instrumentation-mcp==0.52.1; python_version >= '3.10' and python_version < '4.0'",
     "opentelemetry-instrumentation-google-genai==0.6b0",
-    "google-genai==2.0.0",
+    "google-genai==2.3.0",
     "deprecated==1.3.1",
-    "a2a-sdk==0.3.26",
+    "a2a-sdk==1.0.3",
     "email-validator==2.2.0",
-    "pyperclip==1.9.0",
+    "pyperclip==1.11.0",
     "keyring==25.7.0",
     "python-frontmatter==1.1.0",
     "watchfiles==1.1.1",
-    "agent-client-protocol==0.9.0",
-    "jsonschema==4.25.1",
-    "tiktoken==0.12.0",
+    "agent-client-protocol==0.10.0",
+    "jsonschema==4.26.0",
+    "tiktoken==0.13.0",
     "uvloop==0.22.1; platform_system != 'Windows'",
     "multilspy==0.0.15",
     "ruamel.yaml==0.19.1",
-    "huggingface_hub==1.14.0",
+    "huggingface_hub==1.15.0",
     "mslex==1.3.0",
 ]
 

resources/shared/smart_prompt.md

@@ -30,7 +30,7 @@ Use `list_resources` to discover bundled internal resources and attached MCP res
 Use the smart tool to load AgentCards temporarily when you need extra agents.
 Use `create_agent_card` to scaffold a minimal card file quickly.
 Use validate to check AgentCard files before running them.
-Use `attach_resource` when you want to send a prompt with one resource attached.
+Use `attach_media` when you want to send local or provider-fetchable media/document content with the next prompt.
 Use `slash_command` when you need interactive-style `/...` command behavior (for example `/mcp ...`, `/skills ...`, `/cards ...`).
 When calling child-agent tools (`agent__*`), follow each tool's schema and
 parameter descriptions exactly.

src/fast_agent/acp/server/agent_acp_server.py

@@ -663,19 +663,23 @@ async def _initialize_session_state(
 
     async def list_sessions(
         self,
+        additional_directories: list[str] | None = None,
         cursor: str | None = None,
         cwd: str | None = None,
         **kwargs: Any,
     ) -> ListSessionsResponse:
+        _ = additional_directories
         return await self._session_store.list_sessions(cursor=cursor, cwd=cwd, **kwargs)
 
     async def load_session(
         self,
         cwd: str,
         session_id: str,
+        additional_directories: list[str] | None = None,
         mcp_servers: list[HttpMcpServer | SseMcpServer | McpServerStdio] | None = None,
         **kwargs: Any,
     ) -> LoadSessionResponse | None:
+        _ = additional_directories
         return await self._session_store.load_session(
             cwd=cwd,
             session_id=session_id,
@@ -687,9 +691,11 @@ async def resume_session(
         self,
         cwd: str,
         session_id: str,
+        additional_directories: list[str] | None = None,
         mcp_servers: list[HttpMcpServer | SseMcpServer | McpServerStdio] | None = None,
         **kwargs: Any,
     ) -> ResumeSessionResponse:
+        _ = additional_directories
         return await self._session_store.resume_session(
             cwd=cwd,
             session_id=session_id,
@@ -700,6 +706,7 @@ async def resume_session(
     async def new_session(
         self,
         cwd: str,
+        additional_directories: list[str] | None = None,
         mcp_servers: list[HttpMcpServer | SseMcpServer | McpServerStdio] | None = None,
         **kwargs: Any,
     ) -> NewSessionResponse:
@@ -708,6 +715,7 @@ async def new_session(
 
         Creates a new ACP session with its own dedicated agent instance.
         """
+        _ = additional_directories
         request_cwd = self._resolve_request_cwd(
             cwd=cwd,
             request_name="session/new",

src/fast_agent/agents/llm_agent.py

@@ -70,9 +70,7 @@
 
 logger = get_logger(__name__)
 
-DEFAULT_CAPABILITIES = AgentCapabilities(
-    streaming=False, push_notifications=False, state_transition_history=False
-)
+DEFAULT_CAPABILITIES = AgentCapabilities(streaming=False, push_notifications=False)
 
 
 class LlmAgent(LlmDecorator):

src/fast_agent/agents/llm_decorator.py

@@ -29,7 +29,7 @@
     from fast_agent.agents.tool_runner import ToolRunnerHooks
     from fast_agent.hooks.lifecycle_hook_loader import AgentLifecycleHooks
 
-from a2a.types import AgentCard
+from a2a.types import AgentCard, AgentInterface
 from mcp import ListToolsResult, Tool
 from mcp.types import (
     CallToolResult,
@@ -1446,7 +1446,12 @@ async def agent_card(self) -> AgentCard:
             skills=[],
             name=self._name,
             description=self.config.description or self.instruction,
-            url=f"fast-agent://agents/{self._name}/",
+            supported_interfaces=[
+                AgentInterface(
+                    url=f"fast-agent://agents/{self._name}/",
+                    protocol_binding="fast-agent",
+                )
+            ],
             version="0.1",
             capabilities=DEFAULT_CAPABILITIES,
             # TODO -- get these from the _llm

src/fast_agent/agents/mcp_agent.py

@@ -25,7 +25,7 @@
 )
 
 import mcp
-from a2a.types import AgentCard, AgentSkill
+from a2a.types import AgentCard, AgentInterface, AgentSkill
 from mcp.types import (
     CallToolResult,
     ContentBlock,
@@ -489,6 +489,12 @@ def _local_filesystem_runtime(self) -> LocalFilesystemRuntime | None:
                 return fallback
         return None
 
+    def _consume_pending_media_attachments(self) -> list[ContentBlock]:
+        local_runtime = self._local_filesystem_runtime()
+        if local_runtime is None:
+            return []
+        return local_runtime.consume_pending_media_attachments()
+
     @property
     def has_filesystem_read_text_file_tool(self) -> bool:
         """Whether the active filesystem runtime currently exposes read_text_file."""
@@ -648,6 +654,12 @@ def _shell_read_text_file_enabled(self) -> bool:
             return True
         return self._context.config.shell_execution.enable_read_text_file
 
+    def _shell_attach_media_mode(self) -> Literal["auto", "on", "off"]:
+        """Return whether shell-enabled agents should expose local attach_media."""
+        if not self._context or not self._context.config:
+            return "auto"
+        return self._context.config.shell_execution.enable_attach_media
+
     def _resolve_shell_edit_tool_mode(self) -> Literal["write_text_file", "apply_patch", "off"]:
         """Return which shell edit tool should be exposed for the current model/config."""
         default_mode: Literal["write_text_file", "apply_patch"]
@@ -713,16 +725,20 @@ def _maybe_enable_local_filesystem_runtime(self, working_directory: Path | None
         enable_write = edit_mode == "write_text_file"
         enable_apply_patch = edit_mode == "apply_patch"
         enable_edit_file = edit_mode == "write_text_file"
+        enable_attach_media = self._shell_attach_media_mode()
+        model_info = self.llm.model_info if self.llm else None
         local_runtime = self._local_filesystem_runtime()
         if local_runtime is not None:
             if working_directory is not None:
                 local_runtime.set_working_directory(working_directory)
+            local_runtime.set_model_info(model_info)
             local_runtime.set_tool_handler_resolver(self._get_tool_handler)
             local_runtime.set_enabled_tools(
                 enable_read=enable_read,
                 enable_write=enable_write,
                 enable_apply_patch=enable_apply_patch,
                 enable_edit_file=enable_edit_file,
+                enable_attach_media=enable_attach_media,
             )
             return
 
@@ -736,6 +752,8 @@ def _maybe_enable_local_filesystem_runtime(self, working_directory: Path | None
             enable_write=enable_write,
             enable_apply_patch=enable_apply_patch,
             enable_edit_file=enable_edit_file,
+            enable_attach_media=enable_attach_media,
+            model_info=model_info,
             tool_handler_resolver=self._get_tool_handler,
         )
         if self._filesystem_runtime is None:
@@ -752,6 +770,7 @@ def _maybe_enable_local_filesystem_runtime(self, working_directory: Path | None
             write_enabled=enable_write,
             apply_patch_enabled=enable_apply_patch,
             edit_file_enabled=enable_edit_file,
+            attach_media_enabled=enable_attach_media,
         )
 
     def _shell_output_limit_overridden(self) -> bool:
@@ -779,11 +798,13 @@ def _on_llm_attached(self, llm: FastAgentLLMProtocol) -> None:
         local_runtime = self._local_filesystem_runtime()
         if local_runtime is not None:
             edit_mode = self._resolve_shell_edit_tool_mode()
+            local_runtime.set_model_info(llm.model_info)
             local_runtime.set_enabled_tools(
                 enable_read=self._shell_read_text_file_enabled(),
                 enable_write=edit_mode == "write_text_file",
                 enable_apply_patch=edit_mode == "apply_patch",
                 enable_edit_file=edit_mode == "write_text_file",
+                enable_attach_media=self._shell_attach_media_mode(),
             )
 
         if self._shell_runtime is None:
@@ -2032,7 +2053,12 @@ async def agent_card(self) -> AgentCard:
             skills=skills,
             name=self._name,
             description=self.config.description or self.instruction,
-            url=f"fast-agent://agents/{self._name}/",
+            supported_interfaces=[
+                AgentInterface(
+                    url=f"fast-agent://agents/{self._name}/",
+                    protocol_binding="fast-agent",
+                )
+            ],
             version="0.1",
             capabilities=DEFAULT_CAPABILITIES,
             default_input_modes=["text/plain"],

src/fast_agent/agents/smart_agent.py

@@ -1592,7 +1592,8 @@ def _enable_smart_tooling(agent: _SmartToolingAgent) -> None:
         description=(
             "Run subagent tasks from a definition in a file or directory. Subagents are defined with "
             "AgentCards. Use action=`run` to load a subagent and send it a message. Optionally supply "
-            "`mcp_connect` targets. Use action=`validate` to check card file validity without running them"
+            "`mcp_connect` targets. Use action=`validate` to check card file validity without running them. "
+            "Do not pass Agent Skill SKILL.md files here; inspect skills with read_text_file/read_skill."
         ),
     )
     slash_command_tool = build_default_function_tool(

src/fast_agent/agents/tool_agent.py

@@ -7,13 +7,14 @@
 from typing import Any, Callable, Dict, List, Mapping, Sequence
 
 from fastmcp.tools import FunctionTool, ToolResult
-from mcp.types import CallToolResult, ListToolsResult, Tool
+from mcp.types import CallToolResult, ContentBlock, ListToolsResult, Tool
 
 from fast_agent.agents.agent_types import AgentConfig, AgentType
 from fast_agent.agents.llm_agent import LlmAgent
 from fast_agent.agents.tool_runner import ToolRunner, ToolRunnerHooks, _ToolLoopAgent
 from fast_agent.constants import (
     FAST_AGENT_ERROR_CHANNEL,
+    FAST_AGENT_PENDING_MEDIA_ATTACHMENTS,
     FAST_AGENT_TOOL_METADATA,
     HUMAN_INPUT_TOOL_NAME,
     should_parallelize_tool_calls,
@@ -607,6 +608,10 @@ def should_suppress_tools_for_structured_turn(
     def _should_display_user_message(self, message: PromptMessageExtended) -> bool:
         return not message.tool_results
 
+    def _consume_pending_media_attachments(self) -> list[ContentBlock]:
+        """Return pending media blocks to send as the next user input."""
+        return []
+
     # we take care of tool results, so skip displaying them
     def show_user_message(self, message: PromptMessageExtended) -> None:
         if message.tool_results:
@@ -810,8 +815,8 @@ def _finalize_tool_results(
         )
         from fast_agent.mcp.url_elicitation_required import URLElicitationRequiredDisplayPayload
 
-        channels = None
-        content = []
+        channels: dict[str, Sequence[ContentBlock]] | None = None
+        content: list[ContentBlock] = []
         if tool_loop_error:
             content.append(text_content(tool_loop_error))
             channels = {
@@ -846,6 +851,12 @@ def _finalize_tool_results(
                 TextContent(type="text", text=json.dumps(deferred_url_elicitations))
             ]
 
+        pending_media = self._consume_pending_media_attachments()
+        if pending_media:
+            if channels is None:
+                channels = {}
+            channels[FAST_AGENT_PENDING_MEDIA_ATTACHMENTS] = pending_media
+
         return PromptMessageExtended(
             role="user",
             content=content,

src/fast_agent/agents/tool_runner.py

@@ -18,6 +18,7 @@
 from fast_agent.constants import (
     DEFAULT_MAX_ITERATIONS,
     FAST_AGENT_ERROR_CHANNEL,
+    FAST_AGENT_PENDING_MEDIA_ATTACHMENTS,
     FAST_AGENT_SYNTHETIC_FINAL_CHANNEL,
     FAST_AGENT_TIMING,
     FAST_AGENT_USAGE,
@@ -582,12 +583,23 @@ def has_pending_tool_response(self) -> bool:
         return self._pending_tool_request is not None
 
     def _stage_tool_response(self, tool_message: PromptMessageExtended) -> None:
+        staged_messages = [tool_message]
+        channels = tool_message.channels
+        if channels and FAST_AGENT_PENDING_MEDIA_ATTACHMENTS in channels:
+            pending_media = channels[FAST_AGENT_PENDING_MEDIA_ATTACHMENTS]
+            visible_channels = dict(channels)
+            del visible_channels[FAST_AGENT_PENDING_MEDIA_ATTACHMENTS]
+            staged_messages = [
+                tool_message.model_copy(update={"channels": visible_channels or None}),
+                PromptMessageExtended(role="user", content=list(pending_media)),
+            ]
+
         if self._use_history_enabled():
-            self._delta_messages = [tool_message]
+            self._delta_messages = staged_messages
         else:
             if self._last_message is not None:
                 self._delta_messages.append(self._last_message)
-            self._delta_messages.append(tool_message)
+            self._delta_messages.extend(staged_messages)
 
     def _should_start_deferred_structured_finalization(
         self,

src/fast_agent/agents/workflow/router_agent.py

@@ -5,8 +5,10 @@
 by determining the best agent for a request and dispatching to it.
 """
 
+import json
 from typing import TYPE_CHECKING, List, Optional, Tuple, Type
 
+from google.protobuf.json_format import MessageToDict
 from mcp import Tool
 from opentelemetry import trace
 from pydantic import BaseModel
@@ -165,9 +167,15 @@ async def _generate_routing_instruction(
         agent_descriptions = []
         for agent in agents:
             agent_card: AgentCard = await agent.agent_card()
+            card_summary = {
+                "name": agent_card.name,
+                "description": agent_card.description,
+                "skills": [MessageToDict(skill) for skill in agent_card.skills],
+            }
             agent_descriptions.append(
-                agent_card.model_dump_json(
-                    include={"name", "description", "skills"}, exclude_none=True
+                json.dumps(
+                    {key: value for key, value in card_summary.items() if value},
+                    separators=(",", ":"),
                 )
             )