rossum-agent 1.0.0rc3__tar.gz → 1.0.0rc5__tar.gz

{rossum_agent-1.0.0rc3/rossum_agent.egg-info → rossum_agent-1.0.0rc5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rossum-agent
-Version: 1.0.0rc3
+Version: 1.0.0rc5
 Summary: AI agent toolkit for Rossum: document workflows conversationally, debug pipelines automatically, and enable agentic configuration of intelligent document processing.
 Author-email: "Dan Stancl (Rossum AI)" <daniel.stancl@gmail.com>
 License: MIT
@@ -161,6 +161,7 @@ uv sync --extra streamlit  # Streamlit UI only
 | `AWS_DEFAULT_REGION` | No | AWS region (default: `us-east-1`) |
 | `REDIS_HOST` | No | Redis host for chat persistence |
 | `REDIS_PORT` | No | Redis port (default: `6379`) |
+| `ROSSUM_MCP_MODE` | No | MCP mode: `read-only` (default) or `read-write` |
 
 ## Usage
 
@@ -298,6 +299,10 @@ flowchart TB
 
 API docs: `/api/docs` (Swagger) or `/api/redoc`
 
+**MCP Mode:** Chat sessions support mode switching via the `mcp_mode` parameter:
+- Set at chat creation: `POST /api/v1/chats` with `{"mcp_mode": "read-write"}`
+- Override per message: `POST /api/v1/chats/{id}/messages` with `{"content": "...", "mcp_mode": "read-write"}`
+
 </details>
 
 ## License

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/README.md

@@ -80,6 +80,7 @@ uv sync --extra streamlit  # Streamlit UI only
 | `AWS_DEFAULT_REGION` | No | AWS region (default: `us-east-1`) |
 | `REDIS_HOST` | No | Redis host for chat persistence |
 | `REDIS_PORT` | No | Redis port (default: `6379`) |
+| `ROSSUM_MCP_MODE` | No | MCP mode: `read-only` (default) or `read-write` |
 
 ## Usage
 
@@ -217,6 +218,10 @@ flowchart TB
 
 API docs: `/api/docs` (Swagger) or `/api/redoc`
 
+**MCP Mode:** Chat sessions support mode switching via the `mcp_mode` parameter:
+- Set at chat creation: `POST /api/v1/chats` with `{"mcp_mode": "read-write"}`
+- Override per message: `POST /api/v1/chats/{id}/messages` with `{"content": "...", "mcp_mode": "read-write"}`
+
 </details>
 
 ## License

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "rossum-agent"
-version = "1.0.0rc3"
+version = "1.0.0rc5"
 description = "AI agent toolkit for Rossum: document workflows conversationally, debug pipelines automatically, and enable agentic configuration of intelligent document processing."
 readme = "README.md"
 requires-python = ">=3.12"

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/__init__.py

@@ -4,6 +4,6 @@ from __future__ import annotations
 
 from rossum_agent.agent import AgentConfig, AgentStep, RossumAgent, create_agent
 
-__version__ = "1.0.0rc3"
+__version__ = "1.0.0rc5"
 
 __all__ = ["AgentConfig", "AgentStep", "RossumAgent", "create_agent"]

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/agent/core.py

@@ -75,7 +75,6 @@ from rossum_agent.agent.models import (
     ToolResult,
     truncate_content,
 )
-from rossum_agent.agent.request_classifier import RequestScope, classify_request, generate_rejection_response
 from rossum_agent.api.models.schemas import TokenUsageBreakdown
 from rossum_agent.bedrock_client import create_bedrock_client, get_model_id
 from rossum_agent.rossum_mcp_integration import MCPConnection, mcp_tools_to_anthropic_format
@@ -769,32 +768,6 @@ class RossumAgent:
                     text_parts.append(text)
         return " ".join(text_parts)
 
-    def _check_request_scope(self, prompt: UserContent) -> AgentStep | None:
-        """Check if request is in scope, return rejection step if out of scope."""
-        text = self._extract_text_from_prompt(prompt)
-        result = classify_request(self.client, text)
-        self._total_input_tokens += result.input_tokens
-        self._total_output_tokens += result.output_tokens
-        self._main_agent_input_tokens += result.input_tokens
-        self._main_agent_output_tokens += result.output_tokens
-        if result.scope == RequestScope.OUT_OF_SCOPE:
-            rejection = generate_rejection_response(self.client, text)
-            total_input = result.input_tokens + rejection.input_tokens
-            total_output = result.output_tokens + rejection.output_tokens
-            self._total_input_tokens += rejection.input_tokens
-            self._total_output_tokens += rejection.output_tokens
-            self._main_agent_input_tokens += rejection.input_tokens
-            self._main_agent_output_tokens += rejection.output_tokens
-            return AgentStep(
-                step_number=1,
-                final_answer=rejection.response,
-                is_final=True,
-                input_tokens=total_input,
-                output_tokens=total_output,
-                step_type=StepType.FINAL_ANSWER,
-            )
-        return None
-
     def _inject_preload_info(self, prompt: UserContent, preload_result: str) -> UserContent:
         """Inject preload result info into the user prompt."""
         suffix = (
@@ -820,10 +793,6 @@ class RossumAgent:
 
         Rate limiting is handled with exponential backoff and jitter.
         """
-        if rejection := self._check_request_scope(prompt):
-            yield rejection
-            return
-
         loop = asyncio.get_event_loop()
         mcp_mode = get_mcp_mode()
         set_mcp_connection(self.mcp_connection, loop, mcp_mode)

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/api/models/schemas.py

@@ -133,6 +133,10 @@ class MessageRequest(BaseModel):
         description="Optional list of PDF documents (max 5) to include with the message",
     )
     rossum_url: str | None = Field(default=None, description="Optional Rossum app URL for context")
+    mcp_mode: Literal["read-only", "read-write"] | None = Field(
+        default=None,
+        description="MCP mode to use for this message and all subsequent messages. If not specified, uses the chat's current mode.",
+    )
 
 
 class StepEvent(BaseModel):

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/api/routes/messages.py

@@ -157,6 +157,11 @@ async def send_message(
 
     history = chat_data.messages
     mcp_mode = chat_data.metadata.mcp_mode
+    # Use message-level mode if provided, otherwise use chat's mode
+    if message.mcp_mode is not None:
+        mcp_mode = message.mcp_mode
+        # Update chat metadata with new mode for future messages
+        chat_data.metadata.mcp_mode = mcp_mode
     user_prompt = message.content
     images: list[ImageContent] | None = message.images
     documents: list[DocumentContent] | None = message.documents

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/api/services/agent_service.py

@@ -3,7 +3,9 @@
 from __future__ import annotations
 
 import asyncio
+import contextvars
 import logging
+from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, Literal
 
 from rossum_agent.agent.core import RossumAgent, create_agent
@@ -43,6 +45,18 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+@dataclass
+class _RequestContext:
+    """Per-request context for agent execution."""
+
+    output_dir: Path | None = None
+    sub_agent_queue: asyncio.Queue[SubAgentProgressEvent | SubAgentTextEvent] | None = None
+    last_memory: AgentMemory | None = None
+
+
+_request_context: contextvars.ContextVar[_RequestContext] = contextvars.ContextVar("request_context")
+
+
 def convert_sub_agent_progress_to_event(progress: SubAgentProgress) -> SubAgentProgressEvent:
     """Convert a SubAgentProgress to a SubAgentProgressEvent for SSE streaming.
 
@@ -133,28 +147,36 @@ class AgentService:
     """Service for running the Rossum Agent.
 
     Manages MCP connection lifecycle and agent execution for API requests.
+    Uses contextvars for per-request state to support concurrent requests.
     """
 
     def __init__(self) -> None:
         """Initialize agent service."""
-        self._output_dir: Path | None = None
-        self._sub_agent_queue: asyncio.Queue[SubAgentProgressEvent | SubAgentTextEvent] | None = None
-        self._last_memory: AgentMemory | None = None
+
+    def _get_context(self) -> _RequestContext:
+        """Get the current request context, creating if needed."""
+        try:
+            return _request_context.get()
+        except LookupError:
+            ctx = _RequestContext()
+            _request_context.set(ctx)
+            return ctx
 
     @property
     def output_dir(self) -> Path | None:
         """Get the output directory for the current run."""
-        return self._output_dir
+        return self._get_context().output_dir
 
     def _on_sub_agent_progress(self, progress: SubAgentProgress) -> None:
         """Callback for sub-agent progress updates.
 
         Converts the progress to an event and puts it on the queue for streaming.
         """
-        if self._sub_agent_queue is not None:
+        ctx = self._get_context()
+        if ctx.sub_agent_queue is not None:
             event = convert_sub_agent_progress_to_event(progress)
             try:
-                self._sub_agent_queue.put_nowait(event)
+                ctx.sub_agent_queue.put_nowait(event)
             except asyncio.QueueFull:
                 logger.warning("Sub-agent progress queue full, dropping event")
 
@@ -163,10 +185,11 @@ class AgentService:
 
         Converts the text to an event and puts it on the queue for streaming.
         """
-        if self._sub_agent_queue is not None:
+        ctx = self._get_context()
+        if ctx.sub_agent_queue is not None:
             event = SubAgentTextEvent(tool_name=text.tool_name, text=text.text, is_final=text.is_final)
             try:
-                self._sub_agent_queue.put_nowait(event)
+                ctx.sub_agent_queue.put_nowait(event)
             except asyncio.QueueFull:
                 logger.warning("Sub-agent text queue full, dropping event")
 
@@ -196,16 +219,19 @@ class AgentService:
         if documents:
             logger.info(f"Including {len(documents)} documents in the prompt")
 
-        self._output_dir = create_session_output_dir()
-        set_session_output_dir(self._output_dir)
-        set_output_dir(self._output_dir)
+        ctx = _RequestContext()
+        _request_context.set(ctx)
+
+        ctx.output_dir = create_session_output_dir()
+        set_session_output_dir(ctx.output_dir)
+        set_output_dir(ctx.output_dir)
         set_rossum_credentials(rossum_api_base_url, rossum_api_token)
-        logger.info(f"Created session output directory: {self._output_dir}")
+        logger.info(f"Created session output directory: {ctx.output_dir}")
 
         if documents:
             self._save_documents_to_output_dir(documents)
 
-        self._sub_agent_queue = asyncio.Queue(maxsize=100)
+        ctx.sub_agent_queue = asyncio.Queue(maxsize=100)
         set_progress_callback(self._on_sub_agent_progress)
         set_text_callback(self._on_sub_agent_text)
 
@@ -237,9 +263,9 @@ class AgentService:
 
                 try:
                     async for step in agent.run(user_content):
-                        while not self._sub_agent_queue.empty():
+                        while not ctx.sub_agent_queue.empty():
                             try:
-                                sub_event = self._sub_agent_queue.get_nowait()
+                                sub_event = ctx.sub_agent_queue.get_nowait()
                                 yield sub_event
                             except asyncio.QueueEmpty:
                                 break
@@ -251,14 +277,14 @@ class AgentService:
                             total_input_tokens = agent._total_input_tokens
                             total_output_tokens = agent._total_output_tokens
 
-                    while not self._sub_agent_queue.empty():
+                    while not ctx.sub_agent_queue.empty():
                         try:
-                            sub_event = self._sub_agent_queue.get_nowait()
+                            sub_event = ctx.sub_agent_queue.get_nowait()
                             yield sub_event
                         except asyncio.QueueEmpty:
                             break
 
-                    self._last_memory = agent.memory
+                    ctx.last_memory = agent.memory
 
                     yield StreamDoneEvent(
                         total_steps=total_steps,
@@ -281,7 +307,6 @@ class AgentService:
             set_text_callback(None)
             set_output_dir(None)
             set_rossum_credentials(None, None)
-            self._sub_agent_queue = None
 
     def _save_documents_to_output_dir(self, documents: list[DocumentContent]) -> None:
         """Save uploaded documents to the output directory.
@@ -291,12 +316,13 @@ class AgentService:
         """
         import base64  # noqa: PLC0415 - import here to avoid circular import at module level
 
-        if self._output_dir is None:
+        ctx = self._get_context()
+        if ctx.output_dir is None:
             logger.warning("Cannot save documents: output directory not set")
             return
 
         for doc in documents:
-            file_path = self._output_dir / doc.filename
+            file_path = ctx.output_dir / doc.filename
             try:
                 file_data = base64.b64decode(doc.data)
                 file_path.write_bytes(file_data)
@@ -320,6 +346,7 @@ class AgentService:
         if not images and not documents:
             return prompt
 
+        ctx = self._get_context()
         content: list[ImageBlockParam | TextBlockParam] = []
         if images:
             for img in images:
@@ -333,8 +360,8 @@ class AgentService:
                         },
                     }
                 )
-        if documents and self._output_dir:
-            doc_paths = [str(self._output_dir / doc.filename) for doc in documents]
+        if documents and ctx.output_dir:
+            doc_paths = [str(ctx.output_dir / doc.filename) for doc in documents]
             doc_info = "\n".join(f"- {path}" for path in doc_paths)
             content.append({"type": "text", "text": f"[Uploaded documents available for processing:\n{doc_info}]"})
         content.append({"type": "text", "text": prompt})
@@ -416,9 +443,10 @@ class AgentService:
             images: Optional list of images included with the user prompt.
             documents: Optional list of documents included with the user prompt.
         """
-        if self._last_memory is not None:
+        ctx = self._get_context()
+        if ctx.last_memory is not None:
             lean_history: list[dict[str, Any]] = []
-            for step_dict in self._last_memory.to_dict():
+            for step_dict in ctx.last_memory.to_dict():
                 if step_dict.get("type") == "task_step":
                     lean_history.append(step_dict)
                 elif step_dict.get("type") == "memory_step":

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent/prompts/base_prompt.py

@@ -5,7 +5,7 @@ Optimized for Opus 4.5: Goals + constraints, not procedures.
 
 from __future__ import annotations
 
-ROSSUM_EXPERT_INTRO = """You are an expert Rossum platform specialist. Help users understand, document, debug, and configure document processing workflows.
+ROSSUM_EXPERT_INTRO = """You are an expert Rossum platform specialist. Help users understand, document, debug, and configure document processing workflows. Politely redirect requests unrelated to Rossum.
 
 **CRITICAL - Use `search_knowledge_base` before**:
 - Explaining ANY extension/hook behavior (except simple function hooks you can read directly)

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5/rossum_agent.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rossum-agent
-Version: 1.0.0rc3
+Version: 1.0.0rc5
 Summary: AI agent toolkit for Rossum: document workflows conversationally, debug pipelines automatically, and enable agentic configuration of intelligent document processing.
 Author-email: "Dan Stancl (Rossum AI)" <daniel.stancl@gmail.com>
 License: MIT
@@ -161,6 +161,7 @@ uv sync --extra streamlit  # Streamlit UI only
 | `AWS_DEFAULT_REGION` | No | AWS region (default: `us-east-1`) |
 | `REDIS_HOST` | No | Redis host for chat persistence |
 | `REDIS_PORT` | No | Redis port (default: `6379`) |
+| `ROSSUM_MCP_MODE` | No | MCP mode: `read-only` (default) or `read-write` |
 
 ## Usage
 
@@ -298,6 +299,10 @@ flowchart TB
 
 API docs: `/api/docs` (Swagger) or `/api/redoc`
 
+**MCP Mode:** Chat sessions support mode switching via the `mcp_mode` parameter:
+- Set at chat creation: `POST /api/v1/chats` with `{"mcp_mode": "read-write"}`
+- Override per message: `POST /api/v1/chats/{id}/messages` with `{"content": "...", "mcp_mode": "read-write"}`
+
 </details>
 
 ## License

{rossum_agent-1.0.0rc3 → rossum_agent-1.0.0rc5}/rossum_agent.egg-info/SOURCES.txt

@@ -20,7 +20,6 @@ rossum_agent/agent/__init__.py
 rossum_agent/agent/core.py
 rossum_agent/agent/memory.py
 rossum_agent/agent/models.py
-rossum_agent/agent/request_classifier.py
 rossum_agent/agent/skills.py
 rossum_agent/agent/types.py
 rossum_agent/api/__init__.py

rossum_agent-1.0.0rc3/rossum_agent/agent/request_classifier.py

@@ -1,152 +0,0 @@
-"""Lightweight request classifier for filtering out-of-scope requests.
-
-This module provides a fast pre-filter that checks if a user request is within
-the scope of the Rossum platform assistant before engaging the full agent.
-"""
-
-from __future__ import annotations
-
-import logging
-from dataclasses import dataclass
-from enum import Enum
-from typing import TYPE_CHECKING
-
-from rossum_agent.bedrock_client import get_small_model_id
-
-if TYPE_CHECKING:
-    from anthropic import AnthropicBedrock
-
-logger = logging.getLogger(__name__)
-
-CLASSIFIER_PROMPT = """You are a scope classifier for a Rossum document processing platform assistant.
-
-The assistant can help with:
-- Queue, hook, schema, and extension analysis/configuration
-- Debugging document processing issues and errors
-- Investigating hook logs and extension behavior
-- Explaining workflows and automation
-- Writing analysis reports about Rossum configuration issues
-
-IN_SCOPE: Request relates to Rossum PLATFORM operations
-- setting up new organization
-- analyzing/configuring queues, hooks, schemas, extensions
-- debugging errors, investigating logs
-- explaining workflows, analysis
-- generating structured report of customer use-cases on the platform
-- generating formula fields suggestions
-- Also: user asks what the assistant can do, greets assistant
-
-OUT_OF_SCOPE: Request is for DATA analytics - aggregating extracted data, generating charts/plots from document data, summarizing line items/amounts across documents, creating files unrelated to Rossum debugging. Even if it mentions Rossum annotations, if the goal is data aggregation/visualization, it's OUT_OF_SCOPE.
-
-Examples:
-- "Investigate errors with document splitting on queue X" → IN_SCOPE (debugging)
-- "Aggregate line item amounts and generate a bar chart" → OUT_OF_SCOPE (data analytics)
-- "Create a markdown saying hello" → OUT_OF_SCOPE (generic file creation)
-
-Respond with exactly one word: IN_SCOPE or OUT_OF_SCOPE
-
-User request: {message}"""
-
-CLASSIFIER_MAX_TOKENS = 10
-
-
-class RequestScope(Enum):
-    """Classification result for a user request."""
-
-    IN_SCOPE = "IN_SCOPE"
-    OUT_OF_SCOPE = "OUT_OF_SCOPE"
-
-
-@dataclass
-class ClassificationResult:
-    """Result of request classification."""
-
-    scope: RequestScope
-    raw_response: str
-    input_tokens: int = 0
-    output_tokens: int = 0
-
-
-@dataclass
-class RejectionResult:
-    """Result of rejection response generation."""
-
-    response: str
-    input_tokens: int = 0
-    output_tokens: int = 0
-
-
-REJECTION_PROMPT = """You are an expert Rossum platform specialist. The user made a request that is outside your scope.
-
-I can help with:
-- Analyzing and debugging hooks, extensions, and workflows
-- Documenting queue configurations
-- Investigating processing errors
-- Configuring automation
-
-The user asked: {message}
-
-Write a brief, helpful response that:
-1. Politely explains this is outside your Rossum platform expertise
-2. Briefly mentions 2-3 relevant things you CAN help with from the capabilities above
-3. Asks if they have any Rossum-related questions
-
-Keep it concise (3-4 sentences max). Be friendly, not robotic."""
-
-REJECTION_MAX_TOKENS = 300
-
-
-def generate_rejection_response(client: AnthropicBedrock, message: str) -> RejectionResult:
-    """Generate a contextual rejection response for out-of-scope requests."""
-    prompt = REJECTION_PROMPT.format(message=message)
-    try:
-        response = client.messages.create(
-            model=get_small_model_id(), max_tokens=REJECTION_MAX_TOKENS, messages=[{"role": "user", "content": prompt}]
-        )
-        text = response.content[0].text.strip() if response.content else _fallback_response()
-        return RejectionResult(
-            response=text, input_tokens=response.usage.input_tokens, output_tokens=response.usage.output_tokens
-        )
-    except Exception as e:
-        logger.warning(f"Rejection response generation failed: {e}")
-        return RejectionResult(response=_fallback_response())
-
-
-def _fallback_response() -> str:
-    return (
-        "I'm an expert Rossum platform specialist focused on document processing workflows. "
-        "Your request appears to be outside my area of expertise. "
-        "I can help with analyzing hooks, debugging extensions, documenting queue configurations, "
-        "and configuring automation workflows. Do you have any Rossum-related questions?"
-    )
-
-
-def classify_request(client: AnthropicBedrock, message: str) -> ClassificationResult:
-    """Classify whether a user request is within scope.
-
-    Uses a fast, cheap model (Haiku) with minimal tokens to quickly determine if the request should be processed by the main agent.
-    """
-    prompt = CLASSIFIER_PROMPT.format(message=message)
-
-    try:
-        response = client.messages.create(
-            model=get_small_model_id(),
-            max_tokens=CLASSIFIER_MAX_TOKENS,
-            messages=[{"role": "user", "content": prompt}],
-        )
-
-        raw_response = response.content[0].text.strip().upper() if response.content else ""
-
-        scope = RequestScope.OUT_OF_SCOPE if "OUT_OF_SCOPE" in raw_response else RequestScope.IN_SCOPE
-
-        logger.debug(f"Request classified as {scope.value}: {message[:50]}...")
-        return ClassificationResult(
-            scope=scope,
-            raw_response=raw_response,
-            input_tokens=response.usage.input_tokens,
-            output_tokens=response.usage.output_tokens,
-        )
-
-    except Exception as e:
-        logger.warning(f"Classification failed, defaulting to IN_SCOPE: {e}")
-        return ClassificationResult(scope=RequestScope.IN_SCOPE, raw_response=f"error: {e}")