remdb 0.2.6__py3-none-any.whl → 0.3.118__py3-none-any.whl

rem/api/routers/chat/streaming.py

@@ -3,20 +3,29 @@ OpenAI-compatible streaming relay for Pydantic AI agents.
 
 Design Pattern:
 - Uses Pydantic AI's agent.iter() to capture full execution including tool calls
-- Streams tool call events with [Calling: tool_name] markers
-- Streams text content deltas as they arrive
+- Emits rich SSE events: reasoning, tool_call, progress, metadata, text_delta
 - Proper OpenAI SSE format with data: prefix and [DONE] terminator
 - Error handling with graceful degradation
 
-Key Insight 
+Key Insight
 - agent.run_stream() stops after first output, missing tool calls
 - agent.iter() provides complete execution with tool call visibility
-- Use PartStartEvent to detect tool calls
-- Use PartDeltaEvent with TextPartDelta for content streaming
+- Use PartStartEvent to detect tool calls and thinking parts
+- Use PartDeltaEvent with TextPartDelta/ThinkingPartDelta for streaming
+- Use PartEndEvent to detect tool completion
+- Use FunctionToolResultEvent to get tool results
 
-SSE Format:
+SSE Format (OpenAI-compatible):
     data: {"id": "chatcmpl-...", "choices": [{"delta": {"content": "..."}}]}\\n\\n
     data: [DONE]\\n\\n
+
+Extended SSE Format (Custom Events):
+    event: reasoning\\ndata: {"type": "reasoning", "content": "..."}\\n\\n
+    event: tool_call\\ndata: {"type": "tool_call", "tool_name": "...", "status": "started"}\\n\\n
+    event: progress\\ndata: {"type": "progress", "step": 1, "total_steps": 3}\\n\\n
+    event: metadata\\ndata: {"type": "metadata", "confidence": 0.95}\\n\\n
+
+See sse_events.py for the full event type definitions.
 """
 
 import json
@@ -27,9 +36,14 @@ from typing import AsyncGenerator
 from loguru import logger
 from pydantic_ai.agent import Agent
 from pydantic_ai.messages import (
+    FunctionToolResultEvent,
     PartDeltaEvent,
+    PartEndEvent,
     PartStartEvent,
+    TextPart,
     TextPartDelta,
+    ThinkingPart,
+    ThinkingPartDelta,
     ToolCallPart,
 )
 
@@ -38,6 +52,14 @@ from .models import (
     ChatCompletionStreamChoice,
     ChatCompletionStreamResponse,
 )
+from .sse_events import (
+    DoneEvent,
+    MetadataEvent,
+    ProgressEvent,
+    ReasoningEvent,
+    ToolCallEvent,
+    format_sse_event,
+)
 
 
 async def stream_openai_response(
@@ -45,47 +67,94 @@ async def stream_openai_response(
     prompt: str,
     model: str,
     request_id: str | None = None,
+    # Message correlation IDs for metadata
+    message_id: str | None = None,
+    in_reply_to: str | None = None,
+    session_id: str | None = None,
+    # Agent info for metadata
+    agent_schema: str | None = None,
 ) -> AsyncGenerator[str, None]:
     """
-    Stream Pydantic AI agent responses in OpenAI SSE format with tool call events.
+    Stream Pydantic AI agent responses with rich SSE events.
+
+    Emits all SSE event types matching the simulator:
+    - reasoning: Model thinking/chain-of-thought (from ThinkingPart)
+    - tool_call: Tool invocation start/complete (from ToolCallPart, FunctionToolResultEvent)
+    - progress: Step indicators for multi-step execution
+    - text_delta: Streamed content (OpenAI-compatible format)
+    - metadata: Message IDs, model info, performance metrics
+    - done: Stream completion
 
     Design Pattern:
     1. Use agent.iter() for complete execution (not run_stream())
     2. Iterate over nodes to capture model requests and tool executions
-    3. Stream tool call start events as [Calling: tool_name]
-    4. Stream text content deltas as they arrive
-    5. Send final chunk with finish_reason="stop"
-    6. Send OpenAI termination marker [DONE]
+    3. Emit rich SSE events for reasoning, tools, progress
+    4. Stream text content in OpenAI-compatible format
+    5. Send metadata and done events at completion
 
     Args:
         agent: Pydantic AI agent instance
         prompt: User prompt to run
         model: Model name for response metadata
         request_id: Optional request ID (generates UUID if not provided)
+        message_id: Database ID of the assistant message being streamed
+        in_reply_to: Database ID of the user message this responds to
+        session_id: Session ID for conversation correlation
 
     Yields:
-        SSE-formatted strings: "data: {json}\\n\\n"
+        SSE-formatted strings
 
     Example Stream:
-        data: {"id": "chatcmpl-123", "choices": [{"delta": {"role": "assistant", "content": ""}}]}
+        event: progress
+        data: {"type": "progress", "step": 1, "total_steps": 3, "label": "Processing", "status": "in_progress"}
 
-        data: {"id": "chatcmpl-123", "choices": [{"delta": {"content": "[Calling: search]"}}]}
+        event: reasoning
+        data: {"type": "reasoning", "content": "Analyzing the request..."}
+
+        event: tool_call
+        data: {"type": "tool_call", "tool_name": "search", "status": "started", "arguments": {...}}
+
+        event: tool_call
+        data: {"type": "tool_call", "tool_name": "search", "status": "completed", "result": "..."}
 
         data: {"id": "chatcmpl-123", "choices": [{"delta": {"content": "Found 3 results..."}}]}
 
-        data: {"id": "chatcmpl-123", "choices": [{"delta": {}, "finish_reason": "stop"}]}
+        event: metadata
+        data: {"type": "metadata", "message_id": "...", "latency_ms": 1234}
 
-        data: [DONE]
+        event: done
+        data: {"type": "done", "reason": "stop"}
     """
     if request_id is None:
         request_id = f"chatcmpl-{uuid.uuid4().hex[:24]}"
 
     created_at = int(time.time())
+    start_time = time.time()
     is_first_chunk = True
+    reasoning_step = 0
+    current_step = 0
+    total_steps = 3  # Model request, tool execution (optional), final response
+    token_count = 0
+
+    # Track active tool calls for completion events
+    # Maps index -> (tool_name, tool_id) for correlating start/end events
+    active_tool_calls: dict[int, tuple[str, str]] = {}
+    # Queue of tool calls awaiting completion (FIFO for matching)
+    pending_tool_completions: list[tuple[str, str]] = []
+    # Track if metadata was registered via register_metadata tool
+    metadata_registered = False
 
     try:
+        # Emit initial progress event
+        current_step = 1
+        yield format_sse_event(ProgressEvent(
+            step=current_step,
+            total_steps=total_steps,
+            label="Processing request",
+            status="in_progress"
+        ))
+
         # Use agent.iter() to get complete execution with tool calls
-        # run_stream() stops after first output, missing tool calls
         async with agent.iter(prompt) as agent_run:
             async for node in agent_run:
                 # Check if this is a model request node (includes tool calls)
@@ -93,34 +162,163 @@ async def stream_openai_response(
                     # Stream events from model request
                     async with node.stream(agent_run.ctx) as request_stream:
                         async for event in request_stream:
-                            # Tool call start event
+                            # ============================================
+                            # REASONING EVENTS (ThinkingPart)
+                            # ============================================
                             if isinstance(event, PartStartEvent) and isinstance(
+                                event.part, ThinkingPart
+                            ):
+                                reasoning_step += 1
+                                if event.part.content:
+                                    yield format_sse_event(ReasoningEvent(
+                                        content=event.part.content,
+                                        step=reasoning_step
+                                    ))
+
+                            # Reasoning delta (streaming thinking)
+                            elif isinstance(event, PartDeltaEvent) and isinstance(
+                                event.delta, ThinkingPartDelta
+                            ):
+                                if event.delta.content_delta:
+                                    yield format_sse_event(ReasoningEvent(
+                                        content=event.delta.content_delta,
+                                        step=reasoning_step
+                                    ))
+
+                            # ============================================
+                            # TEXT CONTENT START (initial text chunk)
+                            # ============================================
+                            elif isinstance(event, PartStartEvent) and isinstance(
+                                event.part, TextPart
+                            ):
+                                # TextPart may contain initial content that needs to be emitted
+                                if event.part.content:
+                                    content = event.part.content
+                                    token_count += len(content.split())
+
+                                    content_chunk = ChatCompletionStreamResponse(
+                                        id=request_id,
+                                        created=created_at,
+                                        model=model,
+                                        choices=[
+                                            ChatCompletionStreamChoice(
+                                                index=0,
+                                                delta=ChatCompletionMessageDelta(
+                                                    role="assistant" if is_first_chunk else None,
+                                                    content=content,
+                                                ),
+                                                finish_reason=None,
+                                            )
+                                        ],
+                                    )
+                                    is_first_chunk = False
+                                    yield f"data: {content_chunk.model_dump_json()}\n\n"
+
+                            # ============================================
+                            # TOOL CALL START EVENTS
+                            # ============================================
+                            elif isinstance(event, PartStartEvent) and isinstance(
                                 event.part, ToolCallPart
                             ):
-                                logger.info(f"🔧 {event.part.tool_name}")
+                                tool_name = event.part.tool_name
 
-                                tool_call_chunk = ChatCompletionStreamResponse(
-                                    id=request_id,
-                                    created=created_at,
-                                    model=model,
-                                    choices=[
-                                        ChatCompletionStreamChoice(
-                                            index=0,
-                                            delta=ChatCompletionMessageDelta(
-                                                role="assistant" if is_first_chunk else None,
-                                                content=f"[Calling: {event.part.tool_name}]",
-                                            ),
-                                            finish_reason=None,
+                                # Handle final_result specially - it's Pydantic AI's
+                                # internal tool for structured output
+                                if tool_name == "final_result":
+                                    # Extract the structured result and emit as content
+                                    args_dict = None
+                                    if event.part.args is not None:
+                                        if hasattr(event.part.args, 'args_dict'):
+                                            args_dict = event.part.args.args_dict
+                                        elif isinstance(event.part.args, dict):
+                                            args_dict = event.part.args
+
+                                    if args_dict:
+                                        # Emit the structured result as JSON content
+                                        result_json = json.dumps(args_dict, indent=2)
+                                        content_chunk = ChatCompletionStreamResponse(
+                                            id=request_id,
+                                            created=created_at,
+                                            model=model,
+                                            choices=[
+                                                ChatCompletionStreamChoice(
+                                                    index=0,
+                                                    delta=ChatCompletionMessageDelta(
+                                                        role="assistant" if is_first_chunk else None,
+                                                        content=result_json,
+                                                    ),
+                                                    finish_reason=None,
+                                                )
+                                            ],
                                         )
-                                    ],
-                                )
-                                is_first_chunk = False
-                                yield f"data: {tool_call_chunk.model_dump_json()}\n\n"
+                                        is_first_chunk = False
+                                        yield f"data: {content_chunk.model_dump_json()}\n\n"
+                                    continue  # Skip regular tool call handling
+
+                                tool_id = f"call_{uuid.uuid4().hex[:8]}"
+                                active_tool_calls[event.index] = (tool_name, tool_id)
+                                # Queue for completion matching (FIFO)
+                                pending_tool_completions.append((tool_name, tool_id))
+
+                                # Emit tool_call SSE event (started)
+                                # Try to get arguments as dict
+                                args_dict = None
+                                if event.part.args is not None:
+                                    if hasattr(event.part.args, 'args_dict'):
+                                        args_dict = event.part.args.args_dict
+                                    elif isinstance(event.part.args, dict):
+                                        args_dict = event.part.args
+
+                                # Log tool call with key parameters
+                                if args_dict and tool_name == "search_rem":
+                                    query_type = args_dict.get("query_type", "?")
+                                    limit = args_dict.get("limit", 20)
+                                    table = args_dict.get("table", "")
+                                    query_text = args_dict.get("query_text", args_dict.get("entity_key", ""))
+                                    if query_text and len(query_text) > 50:
+                                        query_text = query_text[:50] + "..."
+                                    logger.info(f"🔧 {tool_name} {query_type.upper()} '{query_text}' table={table} limit={limit}")
+                                else:
+                                    logger.info(f"🔧 {tool_name}")
+
+                                yield format_sse_event(ToolCallEvent(
+                                    tool_name=tool_name,
+                                    tool_id=tool_id,
+                                    status="started",
+                                    arguments=args_dict
+                                ))
 
-                            # Text content delta
+                                # Update progress
+                                current_step = 2
+                                total_steps = 4  # Added tool execution step
+                                yield format_sse_event(ProgressEvent(
+                                    step=current_step,
+                                    total_steps=total_steps,
+                                    label=f"Calling {tool_name}",
+                                    status="in_progress"
+                                ))
+
+                            # ============================================
+                            # TOOL CALL COMPLETION (PartEndEvent)
+                            # ============================================
+                            elif isinstance(event, PartEndEvent) and isinstance(
+                                event.part, ToolCallPart
+                            ):
+                                if event.index in active_tool_calls:
+                                    tool_name, tool_id = active_tool_calls[event.index]
+                                    # Note: result comes from FunctionToolResultEvent below
+                                    # For now, mark as completed without result
+                                    del active_tool_calls[event.index]
+
+                            # ============================================
+                            # TEXT CONTENT DELTA
+                            # ============================================
                             elif isinstance(event, PartDeltaEvent) and isinstance(
                                 event.delta, TextPartDelta
                             ):
+                                content = event.delta.content_delta
+                                token_count += len(content.split())  # Rough token estimate
+
                                 content_chunk = ChatCompletionStreamResponse(
                                     id=request_id,
                                     created=created_at,
@@ -130,7 +328,7 @@ async def stream_openai_response(
                                             index=0,
                                             delta=ChatCompletionMessageDelta(
                                                 role="assistant" if is_first_chunk else None,
-                                                content=event.delta.content_delta,
+                                                content=content,
                                             ),
                                             finish_reason=None,
                                         )
@@ -139,16 +337,172 @@ async def stream_openai_response(
                                 is_first_chunk = False
                                 yield f"data: {content_chunk.model_dump_json()}\n\n"
 
-                # Check if this is a tool execution node
+                # ============================================
+                # TOOL EXECUTION NODE
+                # ============================================
                 elif Agent.is_call_tools_node(node):
-                    # Stream tool execution - tools complete here
                     async with node.stream(agent_run.ctx) as tools_stream:
-                        async for event in tools_stream:
-                            # We can log tool completion here if needed
-                            # For now, we already logged the call start above
-                            pass
+                        async for tool_event in tools_stream:
+                            # Tool result event - emit completion
+                            if isinstance(tool_event, FunctionToolResultEvent):
+                                # Get the tool name/id from the pending queue (FIFO)
+                                if pending_tool_completions:
+                                    tool_name, tool_id = pending_tool_completions.pop(0)
+                                else:
+                                    # Fallback if queue is empty (shouldn't happen)
+                                    tool_name = "tool"
+                                    tool_id = f"call_{uuid.uuid4().hex[:8]}"
+
+                                # Check if this is a register_metadata tool result
+                                # It returns a dict with _metadata_event: True marker
+                                result_content = tool_event.result.content if hasattr(tool_event.result, 'content') else tool_event.result
+                                is_metadata_event = False
+
+                                if isinstance(result_content, dict) and result_content.get("_metadata_event"):
+                                    is_metadata_event = True
+                                    metadata_registered = True  # Skip default metadata at end
+                                    # Emit MetadataEvent with registered values
+                                    registered_confidence = result_content.get("confidence")
+                                    registered_sources = result_content.get("sources")
+                                    registered_references = result_content.get("references")
+                                    registered_flags = result_content.get("flags")
+                                    # Risk assessment fields
+                                    registered_risk_level = result_content.get("risk_level")
+                                    registered_risk_score = result_content.get("risk_score")
+                                    registered_risk_reasoning = result_content.get("risk_reasoning")
+                                    registered_recommended_action = result_content.get("recommended_action")
+                                    # Extra fields
+                                    registered_extra = result_content.get("extra")
+
+                                    logger.info(
+                                        f"📊 Metadata registered: confidence={registered_confidence}, "
+                                        f"risk_level={registered_risk_level}, sources={registered_sources}"
+                                    )
+
+                                    # Build extra dict with risk fields and any custom extras
+                                    extra_data = {}
+                                    if registered_risk_level is not None:
+                                        extra_data["risk_level"] = registered_risk_level
+                                    if registered_risk_score is not None:
+                                        extra_data["risk_score"] = registered_risk_score
+                                    if registered_risk_reasoning is not None:
+                                        extra_data["risk_reasoning"] = registered_risk_reasoning
+                                    if registered_recommended_action is not None:
+                                        extra_data["recommended_action"] = registered_recommended_action
+                                    if registered_extra:
+                                        extra_data.update(registered_extra)
+
+                                    # Emit metadata event immediately
+                                    yield format_sse_event(MetadataEvent(
+                                        message_id=message_id,
+                                        in_reply_to=in_reply_to,
+                                        session_id=session_id,
+                                        agent_schema=agent_schema,
+                                        confidence=registered_confidence,
+                                        sources=registered_sources,
+                                        model_version=model,
+                                        flags=registered_flags,
+                                        extra=extra_data if extra_data else None,
+                                        hidden=False,
+                                    ))
+
+                                if not is_metadata_event:
+                                    # Normal tool completion - emit ToolCallEvent
+                                    result_str = str(result_content)
+                                    result_summary = result_str[:200] + "..." if len(result_str) > 200 else result_str
+
+                                    # Log result count for search_rem
+                                    if tool_name == "search_rem" and isinstance(result_content, dict):
+                                        results = result_content.get("results", {})
+                                        # Handle nested result structure: results may be a dict with 'results' list and 'count'
+                                        if isinstance(results, dict):
+                                            count = results.get("count", len(results.get("results", [])))
+                                            query_type = results.get("query_type", "?")
+                                            query_text = results.get("query_text", results.get("key", ""))
+                                            table = results.get("table_name", "")
+                                        elif isinstance(results, list):
+                                            count = len(results)
+                                            query_type = "?"
+                                            query_text = ""
+                                            table = ""
+                                        else:
+                                            count = "?"
+                                            query_type = "?"
+                                            query_text = ""
+                                            table = ""
+                                        status = result_content.get("status", "unknown")
+                                        # Truncate query text for logging
+                                        if query_text and len(str(query_text)) > 40:
+                                            query_text = str(query_text)[:40] + "..."
+                                        logger.info(f"  ↳ {tool_name} {query_type} '{query_text}' table={table} → {count} results")
+
+                                    yield format_sse_event(ToolCallEvent(
+                                        tool_name=tool_name,
+                                        tool_id=tool_id,
+                                        status="completed",
+                                        result=result_summary
+                                    ))
+
+                                # Update progress after tool completion
+                                current_step = 3
+                                yield format_sse_event(ProgressEvent(
+                                    step=current_step,
+                                    total_steps=total_steps,
+                                    label="Generating response",
+                                    status="in_progress"
+                                ))
+
+            # After iteration completes, check for structured result
+            # This handles agents with result_type (structured output)
+            # Skip for plain text output - already streamed via TextPartDelta
+            try:
+                result = agent_run.result
+                if result is not None and hasattr(result, 'output'):
+                    output = result.output
+
+                    # Skip plain string output - already streamed via TextPartDelta
+                    # Non-structured output agents (structured_output: false) return strings
+                    if isinstance(output, str):
+                        logger.debug("Plain text output already streamed via TextPartDelta, skipping final emission")
+                    else:
+                        # Serialize the structured output (Pydantic models)
+                        if hasattr(output, 'model_dump'):
+                            # Pydantic model
+                            result_dict = output.model_dump()
+                        elif hasattr(output, '__dict__'):
+                            result_dict = output.__dict__
+                        else:
+                            # Fallback for unknown types
+                            result_dict = {"result": str(output)}
+
+                        result_json = json.dumps(result_dict, indent=2, default=str)
+                        token_count += len(result_json.split())
+
+                        # Emit structured result as content
+                        result_chunk = ChatCompletionStreamResponse(
+                            id=request_id,
+                            created=created_at,
+                            model=model,
+                            choices=[
+                                ChatCompletionStreamChoice(
+                                    index=0,
+                                    delta=ChatCompletionMessageDelta(
+                                        role="assistant" if is_first_chunk else None,
+                                        content=result_json,
+                                    ),
+                                    finish_reason=None,
+                                )
+                            ],
+                        )
+                        is_first_chunk = False
+                        yield f"data: {result_chunk.model_dump_json()}\n\n"
+            except Exception as e:
+                logger.debug(f"No structured result available: {e}")
 
-        # Final chunk with finish_reason
+        # Calculate latency
+        latency_ms = int((time.time() - start_time) * 1000)
+
+        # Final OpenAI chunk with finish_reason
         final_chunk = ChatCompletionStreamResponse(
             id=request_id,
             created=created_at,
@@ -163,7 +517,32 @@ async def stream_openai_response(
         )
         yield f"data: {final_chunk.model_dump_json()}\n\n"
 
-        # OpenAI termination marker
+        # Emit metadata event only if not already registered via register_metadata tool
+        if not metadata_registered:
+            yield format_sse_event(MetadataEvent(
+                message_id=message_id,
+                in_reply_to=in_reply_to,
+                session_id=session_id,
+                agent_schema=agent_schema,
+                confidence=1.0,  # Default to 100% confidence
+                model_version=model,
+                latency_ms=latency_ms,
+                token_count=token_count,
+            ))
+
+        # Mark all progress complete
+        for step in range(1, total_steps + 1):
+            yield format_sse_event(ProgressEvent(
+                step=step,
+                total_steps=total_steps,
+                label="Complete" if step == total_steps else f"Step {step}",
+                status="completed"
+            ))
+
+        # Emit done event
+        yield format_sse_event(DoneEvent(reason="stop"))
+
+        # OpenAI termination marker (for compatibility)
         yield "data: [DONE]\n\n"
 
     except Exception as e:
@@ -182,4 +561,188 @@ async def stream_openai_response(
             }
         }
         yield f"data: {json.dumps(error_data)}\n\n"
+
+        # Emit done event with error reason
+        yield format_sse_event(DoneEvent(reason="error"))
         yield "data: [DONE]\n\n"
+
+
+async def stream_simulator_response(
+    prompt: str,
+    model: str = "simulator-v1.0.0",
+    request_id: str | None = None,
+    delay_ms: int = 50,
+    include_reasoning: bool = True,
+    include_progress: bool = True,
+    include_tool_calls: bool = True,
+    include_actions: bool = True,
+    include_metadata: bool = True,
+    # Message correlation IDs
+    message_id: str | None = None,
+    in_reply_to: str | None = None,
+    session_id: str | None = None,
+) -> AsyncGenerator[str, None]:
+    """
+    Stream SSE simulator events for testing and demonstration.
+
+    This function wraps the SSE simulator to produce formatted SSE strings
+    ready for HTTP streaming. No LLM calls are made.
+
+    The simulator produces a rich sequence of events:
+    1. Reasoning events (model thinking)
+    2. Progress events (step indicators)
+    3. Tool call events (simulated tool usage)
+    4. Text delta events (streamed content)
+    5. Metadata events (confidence, sources, message IDs)
+    6. Action request events (user interaction)
+    7. Done event
+
+    Args:
+        prompt: User prompt (passed to simulator)
+        model: Model name for metadata
+        request_id: Optional request ID
+        delay_ms: Delay between events in milliseconds
+        include_reasoning: Whether to emit reasoning events
+        include_progress: Whether to emit progress events
+        include_tool_calls: Whether to emit tool call events
+        include_actions: Whether to emit action request at end
+        include_metadata: Whether to emit metadata event
+        message_id: Database ID of the assistant message being streamed
+        in_reply_to: Database ID of the user message this responds to
+        session_id: Session ID for conversation correlation
+
+    Yields:
+        SSE-formatted strings ready for HTTP response
+
+    Example:
+        ```python
+        from starlette.responses import StreamingResponse
+
+        async def simulator_endpoint():
+            return StreamingResponse(
+                stream_simulator_response("demo"),
+                media_type="text/event-stream"
+            )
+        ```
+    """
+    from rem.agentic.agents.sse_simulator import stream_simulator_events
+
+    # Simulator now yields SSE-formatted strings directly (OpenAI-compatible)
+    async for sse_string in stream_simulator_events(
+        prompt=prompt,
+        delay_ms=delay_ms,
+        include_reasoning=include_reasoning,
+        include_progress=include_progress,
+        include_tool_calls=include_tool_calls,
+        include_actions=include_actions,
+        include_metadata=include_metadata,
+        # Pass message correlation IDs
+        message_id=message_id,
+        in_reply_to=in_reply_to,
+        session_id=session_id,
+        model=model,
+    ):
+        yield sse_string
+
+
+async def stream_minimal_simulator(
+    content: str = "Hello from the simulator!",
+    delay_ms: int = 30,
+) -> AsyncGenerator[str, None]:
+    """
+    Stream minimal simulator output (text + done only).
+
+    Useful for simple testing without the full event sequence.
+
+    Args:
+        content: Text content to stream
+        delay_ms: Delay between chunks
+
+    Yields:
+        SSE-formatted strings
+    """
+    from rem.agentic.agents.sse_simulator import stream_minimal_demo
+
+    # Simulator now yields SSE-formatted strings directly (OpenAI-compatible)
+    async for sse_string in stream_minimal_demo(content=content, delay_ms=delay_ms):
+        yield sse_string
+
+
+async def stream_openai_response_with_save(
+    agent: Agent,
+    prompt: str,
+    model: str,
+    request_id: str | None = None,
+    agent_schema: str | None = None,
+    session_id: str | None = None,
+    user_id: str | None = None,
+) -> AsyncGenerator[str, None]:
+    """
+    Wrapper around stream_openai_response that saves the assistant response after streaming.
+
+    This accumulates all text content during streaming and saves it to the database
+    after the stream completes.
+
+    Args:
+        agent: Pydantic AI agent instance
+        prompt: User prompt
+        model: Model name
+        request_id: Optional request ID
+        agent_schema: Agent schema name
+        session_id: Session ID for message storage
+        user_id: User ID for message storage
+
+    Yields:
+        SSE-formatted strings
+    """
+    from ....utils.date_utils import utc_now, to_iso
+    from ....services.session import SessionMessageStore
+    from ....settings import settings
+
+    # Accumulate content during streaming
+    accumulated_content = []
+
+    async for chunk in stream_openai_response(
+        agent=agent,
+        prompt=prompt,
+        model=model,
+        request_id=request_id,
+        agent_schema=agent_schema,
+        session_id=session_id,
+    ):
+        yield chunk
+
+        # Extract text content from OpenAI-format chunks
+        # Format: data: {"choices": [{"delta": {"content": "..."}}]}
+        if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
+            try:
+                data_str = chunk[6:].strip()  # Remove "data: " prefix
+                if data_str:
+                    data = json.loads(data_str)
+                    if "choices" in data and data["choices"]:
+                        delta = data["choices"][0].get("delta", {})
+                        content = delta.get("content")
+                        if content:
+                            accumulated_content.append(content)
+            except (json.JSONDecodeError, KeyError, IndexError):
+                pass  # Skip non-JSON or malformed chunks
+
+    # After streaming completes, save the assistant response
+    if settings.postgres.enabled and session_id and accumulated_content:
+        full_content = "".join(accumulated_content)
+        assistant_message = {
+            "role": "assistant",
+            "content": full_content,
+            "timestamp": to_iso(utc_now()),
+        }
+        try:
+            store = SessionMessageStore(user_id=user_id or settings.test.effective_user_id)
+            await store.store_session_messages(
+                session_id=session_id,
+                messages=[assistant_message],
+                user_id=user_id,
+                compress=True,  # Compress long assistant responses
+            )
+            logger.debug(f"Saved assistant response to session {session_id} ({len(full_content)} chars)")
+        except Exception as e:
+            logger.error(f"Failed to save assistant response: {e}", exc_info=True)