[router] harmony responses api streaming support (#12395)

sgl-router/src/routers/grpc/harmony/mod.rs

@@ -43,7 +43,9 @@ pub use builder::HarmonyBuilder;
 pub use detector::HarmonyDetector;
 pub use parser::HarmonyParserAdapter;
 pub use processor::{HarmonyResponseProcessor, ResponsesIterationResult};
-pub use responses::{serve_harmony_responses, HarmonyResponsesContext};
+pub use responses::{
+    serve_harmony_responses, serve_harmony_responses_stream, HarmonyResponsesContext,
+};
 pub use stages::{
     HarmonyPreparationStage, HarmonyRequestBuildingStage, HarmonyResponseProcessingStage,
 };

sgl-router/src/routers/grpc/harmony/parser.rs

@@ -103,7 +103,7 @@ impl HarmonyParserAdapter {
     /// # Returns
     ///
     /// Tuple of (analysis, commentary, final_text)
-    fn parse_messages(
+    pub fn parse_messages(
         messages: &[openai_harmony::chat::Message],
     ) -> (Option<String>, Option<Vec<ToolCall>>, String) {
         let mut analysis = None;
@@ -260,6 +260,51 @@ impl HarmonyParserAdapter {
         self.parser.messages().to_vec()
     }
 
+    /// Extract incomplete commentary content from parser state
+    ///
+    /// When the stream ends, there may be incomplete commentary content in the parser
+    /// that hasn't been finalized into a completed message. This method extracts
+    /// such content and converts it to tool calls.
+    ///
+    /// # Returns
+    ///
+    /// Optional vector of ToolCall if incomplete commentary is found
+    pub fn extract_incomplete_commentary(&self) -> Option<Vec<ToolCall>> {
+        // Check if current channel is commentary
+        let current_channel = self.parser.current_channel();
+        if current_channel.as_deref() != Some("commentary") {
+            return None;
+        }
+
+        // Get current recipient (should be "functions.{name}")
+        let recipient = self.parser.current_recipient()?;
+        if !recipient.starts_with("functions.") {
+            return None;
+        }
+
+        // Get current incomplete content
+        let content = self.parser.current_content().ok()?;
+        if content.is_empty() {
+            return None;
+        }
+
+        // Extract function name from recipient
+        let function_name = recipient.strip_prefix("functions.").unwrap();
+
+        // Create tool call from incomplete content
+        let call_id = format!("call_{}", Uuid::new_v4());
+        let tool_call = ToolCall {
+            id: call_id,
+            tool_type: "function".to_string(),
+            function: FunctionCallResponse {
+                name: function_name.to_string(),
+                arguments: Some(content),
+            },
+        };
+
+        Some(vec![tool_call])
+    }
+
     /// Parse streaming chunk
     ///
     /// Parses incremental token IDs and returns a delta with any new content

sgl-router/src/routers/grpc/harmony/responses.rs

@@ -36,10 +36,17 @@
 //! See `/Users/simolin/workspace/sglang/.claude/docs/harmony_pipeline/tool_loop_design.md`
 //! for complete architecture, rationale, and implementation details.
 
-use std::sync::Arc;
+use std::{
+    sync::Arc,
+    time::{SystemTime, UNIX_EPOCH},
+};
 
-use axum::response::Response;
+use axum::{body::Body, http::StatusCode, response::Response};
 use serde_json::Value as JsonValue;
+use tokio::sync::mpsc;
+use tokio_stream::wrappers::UnboundedReceiverStream;
+use tracing::{debug, warn};
+use uuid::Uuid;
 
 use crate::{
     data_connector::{ResponseId, ResponseStorage},
@@ -47,13 +54,19 @@ use crate::{
     protocols::{
         common::{Function, ToolCall},
         responses::{
-            ResponseInput, ResponseInputOutputItem, ResponseTool, ResponsesRequest,
-            ResponsesResponse, StringOrContentParts,
+            ResponseInput, ResponseInputOutputItem, ResponseTool, ResponseToolType,
+            ResponsesRequest, ResponsesResponse, StringOrContentParts,
         },
     },
-    routers::grpc::{
-        context::SharedComponents, harmony::processor::ResponsesIterationResult,
-        pipeline::RequestPipeline, utils,
+    routers::{
+        grpc::{
+            context::SharedComponents,
+            harmony::processor::ResponsesIterationResult,
+            pipeline::RequestPipeline,
+            responses::streaming::{OutputItemType, ResponseStreamEventEmitter},
+            utils,
+        },
+        openai::mcp::ensure_request_mcp_client,
     },
 };
 
@@ -93,7 +106,7 @@ struct McpCallTracking {
 }
 
 impl McpCallTracking {
-    fn new(server_label: String) -> Self {
+    pub fn new(server_label: String) -> Self {
         Self {
             server_label,
             tool_calls: Vec::new(),
@@ -143,7 +156,7 @@ pub struct HarmonyResponsesContext {
     pub response_storage: Arc<dyn ResponseStorage>,
 
     /// Optional streaming sender (for future streaming support)
-    pub stream_tx: Option<tokio::sync::mpsc::UnboundedSender<Result<String, String>>>,
+    pub stream_tx: Option<mpsc::UnboundedSender<Result<String, String>>>,
 }
 
 impl HarmonyResponsesContext {
@@ -169,7 +182,7 @@ impl HarmonyResponsesContext {
         components: Arc<SharedComponents>,
         mcp_manager: Arc<McpManager>,
         response_storage: Arc<dyn ResponseStorage>,
-        stream_tx: tokio::sync::mpsc::UnboundedSender<Result<String, String>>,
+        stream_tx: mpsc::UnboundedSender<Result<String, String>>,
     ) -> Self {
         Self {
             pipeline,
@@ -226,12 +239,6 @@ pub async fn serve_harmony_responses(
     let mut current_request = load_previous_messages(ctx, request).await?;
     let mut iteration_count = 0;
 
-    // Check if request has MCP tools - if so, ensure dynamic client is registered
-    // and add static MCP tools to the request
-    use crate::{
-        protocols::responses::ResponseToolType, routers::openai::mcp::ensure_request_mcp_client,
-    };
-
     let has_mcp_tools = current_request
         .tools
         .as_ref()
@@ -265,7 +272,7 @@ pub async fn serve_harmony_responses(
             all_tools.extend(mcp_response_tools);
             current_request.tools = Some(all_tools);
 
-            tracing::debug!(
+            debug!(
                 mcp_tool_count = mcp_tools.len(),
                 total_tool_count = current_request.tools.as_ref().map(|t| t.len()).unwrap_or(0),
                 "Request has MCP tools - added static MCP tools to Harmony Responses request"
@@ -284,7 +291,7 @@ pub async fn serve_harmony_responses(
             )));
         }
 
-        tracing::debug!(
+        debug!(
             iteration = iteration_count,
             "Harmony Responses serving iteration"
         );
@@ -308,7 +315,7 @@ pub async fn serve_harmony_responses(
                 analysis,
                 partial_text,
             } => {
-                tracing::debug!(
+                debug!(
                     tool_call_count = tool_calls.len(),
                     has_analysis = analysis.is_some(),
                     partial_text_len = partial_text.len(),
@@ -347,7 +354,7 @@ pub async fn serve_harmony_responses(
                 mut response,
                 usage,
             } => {
-                tracing::debug!(
+                debug!(
                     output_items = response.output.len(),
                     input_tokens = usage.prompt_tokens,
                     output_tokens = usage.completion_tokens,
@@ -360,7 +367,7 @@ pub async fn serve_harmony_responses(
                 if let Some(tracking) = mcp_tracking {
                     inject_mcp_metadata(&mut response, &tracking, &ctx.mcp_manager);
 
-                    tracing::debug!(
+                    debug!(
                         mcp_calls = tracking.total_calls(),
                         output_items_after = response.output.len(),
                         "Injected MCP metadata into final response"
@@ -375,6 +382,327 @@ pub async fn serve_harmony_responses(
     }
 }
 
+/// Serve Harmony Responses API with streaming (SSE)
+///
+/// This is the streaming equivalent of `serve_harmony_responses()`.
+/// Emits SSE events for lifecycle, MCP list_tools, and per-iteration streaming.
+///
+/// # Architecture
+///
+/// - Emits `response.created` and `response.in_progress` at start
+/// - Emits `mcp_list_tools` events on first iteration (if MCP tools available)
+/// - Loops through tool execution iterations (max 10)
+/// - Calls `streaming::process_responses_iteration_stream()` for per-iteration events
+/// - Emits `response.completed` at end
+/// - Handles errors with `response.failed`
+///
+/// # Arguments
+///
+/// * `ctx` - Harmony responses context with pipeline and dependencies
+/// * `request` - Responses API request
+///
+/// # Returns
+///
+/// SSE stream response with proper headers
+pub async fn serve_harmony_responses_stream(
+    ctx: &HarmonyResponsesContext,
+    request: ResponsesRequest,
+) -> Response {
+    // Load previous conversation history if previous_response_id is set
+    let mut current_request = match load_previous_messages(ctx, request).await {
+        Ok(req) => req,
+        Err(err_response) => return err_response,
+    };
+
+    use std::io;
+
+    use bytes::Bytes;
+
+    // Create SSE channel
+    let (tx, rx) = mpsc::unbounded_channel();
+    let stream = UnboundedReceiverStream::new(rx);
+
+    // Create response event emitter
+    let response_id = format!("resp_{}", Uuid::new_v4());
+    let model = current_request.model.clone();
+    let created_at = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .unwrap()
+        .as_secs();
+    let mut emitter = ResponseStreamEventEmitter::new(response_id.clone(), model, created_at);
+
+    // Clone context for spawned task
+    let ctx_clone = ctx.clone();
+
+    // Spawn async task to handle streaming
+    tokio::spawn(async move {
+        let ctx = &ctx_clone;
+
+        // Clone response_id for closure to avoid borrow conflicts
+        let response_id_for_error = response_id.clone();
+
+        // Helper to emit error and return
+        let emit_error = |tx: &mpsc::UnboundedSender<Result<Bytes, io::Error>>, error_msg: &str| {
+            // Create error event manually since emit_failed doesn't exist
+            let event = serde_json::json!({
+                "type": "response.failed",
+                "response_id": response_id_for_error,
+                "error": {
+                    "message": error_msg,
+                    "type": "internal_error"
+                }
+            });
+            let sse_data = format!("data: {}\n\n", serde_json::to_string(&event).unwrap());
+            let _ = tx.send(Ok(Bytes::from(sse_data)));
+        };
+
+        // Emit initial response.created and response.in_progress events
+        let event = emitter.emit_created();
+        if emitter.send_event(&event, &tx).is_err() {
+            return;
+        }
+        let event = emitter.emit_in_progress();
+        if emitter.send_event(&event, &tx).is_err() {
+            return;
+        }
+
+        // Check if request has MCP tools
+        let has_mcp_tools = current_request
+            .tools
+            .as_ref()
+            .map(|tools| {
+                tools
+                    .iter()
+                    .any(|t| matches!(t.r#type, ResponseToolType::Mcp))
+            })
+            .unwrap_or(false);
+
+        // Initialize MCP call tracking
+        let mut mcp_tracking = if has_mcp_tools {
+            Some(McpCallTracking::new("sglang-mcp".to_string()))
+        } else {
+            None
+        };
+
+        // Setup MCP tools if needed
+        if has_mcp_tools {
+            // Ensure dynamic MCP client is registered
+            if let Some(tools) = &current_request.tools {
+                ensure_request_mcp_client(&ctx.mcp_manager, tools).await;
+            }
+
+            // Add static MCP tools from inventory
+            let mcp_tools = ctx.mcp_manager.list_tools();
+            if !mcp_tools.is_empty() {
+                let mcp_response_tools = convert_mcp_tools_to_response_tools(&mcp_tools);
+                let mut all_tools = current_request.tools.clone().unwrap_or_default();
+                all_tools.extend(mcp_response_tools);
+                current_request.tools = Some(all_tools);
+
+                debug!(
+                    mcp_tool_count = mcp_tools.len(),
+                    total_tool_count = current_request.tools.as_ref().map(|t| t.len()).unwrap_or(0),
+                    "Added static MCP tools to Harmony Responses streaming request"
+                );
+            }
+        }
+
+        // Emit mcp_list_tools on first iteration (only if MCP tools available)
+        if has_mcp_tools {
+            let mcp_tools = ctx.mcp_manager.list_tools();
+            let (output_index, item_id) =
+                emitter.allocate_output_index(OutputItemType::McpListTools);
+
+            // Build tools list for item structure
+            let tool_items: Vec<_> = mcp_tools
+                .iter()
+                .map(|t| {
+                    use serde_json::{json, Value};
+                    json!({
+                        "name": t.name,
+                        "description": t.description,
+                        "input_schema": Value::Object((*t.input_schema).clone())
+                    })
+                })
+                .collect();
+
+            // Emit output_item.added
+            let item = serde_json::json!({
+                "id": item_id,
+                "type": "mcp_list_tools",
+                "server_label": "sglang-mcp",
+                "status": "in_progress",
+                "tools": []
+            });
+            let event = emitter.emit_output_item_added(output_index, &item);
+            if emitter.send_event(&event, &tx).is_err() {
+                return;
+            }
+
+            // Emit mcp_list_tools.in_progress
+            let event = emitter.emit_mcp_list_tools_in_progress(output_index);
+            if emitter.send_event(&event, &tx).is_err() {
+                return;
+            }
+
+            // Emit mcp_list_tools.completed
+            let event = emitter.emit_mcp_list_tools_completed(output_index, &mcp_tools);
+            if emitter.send_event(&event, &tx).is_err() {
+                return;
+            }
+
+            // Emit output_item.done
+            let item_done = serde_json::json!({
+                "id": item_id,
+                "type": "mcp_list_tools",
+                "server_label": "sglang-mcp",
+                "status": "completed",
+                "tools": tool_items
+            });
+            let event = emitter.emit_output_item_done(output_index, &item_done);
+            if emitter.send_event(&event, &tx).is_err() {
+                return;
+            }
+
+            emitter.complete_output_item(output_index);
+
+            debug!(
+                tool_count = mcp_tools.len(),
+                "Emitted mcp_list_tools on first iteration"
+            );
+        }
+
+        // Tool loop (max 10 iterations)
+        let mut iteration_count = 0;
+        loop {
+            iteration_count += 1;
+
+            // Safety check: prevent infinite loops
+            if iteration_count > MAX_TOOL_ITERATIONS {
+                let error_msg =
+                    format!("Maximum tool iterations ({}) exceeded", MAX_TOOL_ITERATIONS);
+                emit_error(&tx, &error_msg);
+                return;
+            }
+
+            debug!(
+                iteration = iteration_count,
+                "Harmony Responses streaming iteration"
+            );
+
+            // Execute through pipeline and get raw stream
+            let execution_result = match ctx
+                .pipeline
+                .execute_harmony_responses_streaming(&current_request, ctx)
+                .await
+            {
+                Ok(result) => result,
+                Err(err_response) => {
+                    let error_msg = format!("Pipeline execution failed: {:?}", err_response);
+                    emit_error(&tx, &error_msg);
+                    return;
+                }
+            };
+
+            // Process stream with token-level streaming using HarmonyStreamingProcessor
+            let iteration_result = match super::streaming::HarmonyStreamingProcessor::process_responses_iteration_stream(
+                execution_result,
+                &mut emitter,
+                &tx,
+            )
+            .await
+            {
+                Ok(result) => result,
+                Err(err_msg) => {
+                    emit_error(&tx, &err_msg);
+                    return;
+                }
+            };
+
+            // Handle iteration result (tool calls or completion)
+            match iteration_result {
+                ResponsesIterationResult::ToolCallsFound {
+                    tool_calls,
+                    analysis,
+                    partial_text,
+                } => {
+                    debug!(
+                        tool_call_count = tool_calls.len(),
+                        has_analysis = analysis.is_some(),
+                        partial_text_len = partial_text.len(),
+                        "Tool calls found in commentary channel"
+                    );
+
+                    // Execute MCP tools
+                    let tool_results = if let Some(ref mut tracking) = mcp_tracking {
+                        match execute_mcp_tools(&ctx.mcp_manager, &tool_calls, tracking).await {
+                            Ok(results) => results,
+                            Err(err_response) => {
+                                let error_msg =
+                                    format!("MCP tool execution failed: {:?}", err_response);
+                                emit_error(&tx, &error_msg);
+                                return;
+                            }
+                        }
+                    } else {
+                        let error_msg = "Tool calls found but MCP tracking not initialized";
+                        emit_error(&tx, error_msg);
+                        return;
+                    };
+
+                    // Build next request with appended history
+                    current_request = match build_next_request_with_tools(
+                        current_request,
+                        tool_calls,
+                        tool_results,
+                        analysis,
+                        partial_text,
+                    ) {
+                        Ok(req) => req,
+                        Err(e) => {
+                            let error_msg = format!("Failed to build next request: {:?}", e);
+                            emit_error(&tx, &error_msg);
+                            return;
+                        }
+                    };
+
+                    // Continue loop
+                }
+                ResponsesIterationResult::Completed { response, usage } => {
+                    debug!(
+                        output_items = response.output.len(),
+                        input_tokens = usage.prompt_tokens,
+                        output_tokens = usage.completion_tokens,
+                        "Harmony Responses streaming completed - no more tool calls"
+                    );
+
+                    // Emit response.completed with usage
+                    let usage_json = serde_json::json!({
+                        "prompt_tokens": usage.prompt_tokens,
+                        "completion_tokens": usage.completion_tokens,
+                        "total_tokens": usage.total_tokens,
+                    });
+                    let event = emitter.emit_completed(Some(&usage_json));
+                    emitter.send_event_best_effort(&event, &tx);
+
+                    // Close channel
+                    drop(tx);
+                    return;
+                }
+            }
+        }
+    });
+
+    // Return SSE stream response
+    Response::builder()
+        .status(StatusCode::OK)
+        .header("Content-Type", "text/event-stream")
+        .header("Cache-Control", "no-cache")
+        .header("Connection", "keep-alive")
+        .body(Body::from_stream(stream))
+        .unwrap()
+}
+
 /// Execute MCP tools and collect results
 ///
 /// Executes each tool call sequentially via the MCP manager.
@@ -397,7 +725,7 @@ async fn execute_mcp_tools(
     let mut results = Vec::new();
 
     for tool_call in tool_calls {
-        tracing::debug!(
+        debug!(
             tool_name = %tool_call.function.name,
             call_id = %tool_call.id,
             "Executing MCP tool"
@@ -425,7 +753,7 @@ async fn execute_mcp_tools(
             .await
         {
             Ok(mcp_result) => {
-                tracing::debug!(
+                debug!(
                     tool_name = %tool_call.function.name,
                     call_id = %tool_call.id,
                     "Tool execution succeeded"
@@ -468,7 +796,7 @@ async fn execute_mcp_tools(
                 });
             }
             Err(e) => {
-                tracing::warn!(
+                warn!(
                     tool_name = %tool_call.function.name,
                     call_id = %tool_call.id,
                     error = %e,
@@ -656,7 +984,7 @@ struct ToolResult {
 /// # Returns
 ///
 /// Vector of ResponseTool entries in MCP format
-fn convert_mcp_tools_to_response_tools(mcp_tools: &[crate::mcp::Tool]) -> Vec<ResponseTool> {
+pub fn convert_mcp_tools_to_response_tools(mcp_tools: &[crate::mcp::Tool]) -> Vec<ResponseTool> {
     use serde_json::Value;
 
     use crate::protocols::responses::ResponseToolType;
@@ -800,11 +1128,9 @@ async fn load_previous_messages(
                 .filter_map(|item| {
                     serde_json::from_value::<ResponseInputOutputItem>(item.clone())
                         .map_err(|e| {
-                            tracing::warn!(
+                            warn!(
                                 "Failed to deserialize stored {} item: {}. Item: {}",
-                                item_type,
-                                e,
-                                item
+                                item_type, e, item
                             );
                         })
                         .ok()
@@ -817,7 +1143,7 @@ async fn load_previous_messages(
         history_items.extend(deserialize_items(&stored.output, "output"));
     }
 
-    tracing::debug!(
+    debug!(
         previous_response_id = %prev_id_str,
         history_items_count = history_items.len(),
         "Loaded conversation history from previous response"
@@ -851,29 +1177,3 @@ async fn load_previous_messages(
 
     Ok(modified_request)
 }
-
-// TODO: Implement streaming support
-// /// Emit intermediate streaming chunks for analysis and partial text
-// ///
-// /// Emits SSE chunks for Responses API streaming:
-// /// - Reasoning chunks for analysis channel
-// /// - Message chunks for partial text from final channel
-// ///
-// /// # Arguments
-// ///
-// /// * `tx` - Streaming sender
-// /// * `analysis` - Analysis channel content
-// /// * `partial_text` - Final channel content
-// /// * `iteration` - Current iteration number
-// async fn emit_intermediate_chunks(
-//     tx: &tokio::sync::mpsc::UnboundedSender<Result<String, String>>,
-//     analysis: &Option<String>,
-//     partial_text: &str,
-//     iteration: usize,
-// ) -> Result<(), Response> {
-//     // TODO: Implement streaming emission
-//     // - Emit reasoning chunks for analysis
-//     // - Emit message chunks for partial_text
-//     // - Follow OpenAI Responses streaming format (14 SSE event types)
-//     Ok(())
-// }

sgl-router/src/routers/grpc/harmony/stages/response_processing.rs

@@ -40,25 +40,26 @@ impl Default for HarmonyResponseProcessingStage {
 #[async_trait]
 impl PipelineStage for HarmonyResponseProcessingStage {
     async fn execute(&self, ctx: &mut RequestContext) -> Result<Option<Response>, Response> {
-        // Get execution result (output tokens from model)
-        let execution_result = ctx
-            .state
-            .response
-            .execution_result
-            .take()
-            .ok_or_else(|| utils::internal_error_static("No execution result"))?;
-
         let is_streaming = ctx.is_streaming();
-        let dispatch = ctx
-            .state
-            .dispatch
-            .as_ref()
-            .cloned()
-            .ok_or_else(|| utils::internal_error_static("Dispatch metadata not set"))?;
 
         // Check request type to determine which processor method to call
         match &ctx.input.request_type {
             RequestType::Chat(_) => {
+                // Get execution result (output tokens from model)
+                let execution_result = ctx
+                    .state
+                    .response
+                    .execution_result
+                    .take()
+                    .ok_or_else(|| utils::internal_error_static("No execution result"))?;
+
+                let dispatch = ctx
+                    .state
+                    .dispatch
+                    .as_ref()
+                    .cloned()
+                    .ok_or_else(|| utils::internal_error_static("Dispatch metadata not set"))?;
+
                 // For streaming, delegate to streaming processor and return SSE response
                 if is_streaming {
                     return Ok(Some(
@@ -83,14 +84,28 @@ impl PipelineStage for HarmonyResponseProcessingStage {
                 Ok(None)
             }
             RequestType::Responses(_) => {
-                // For Responses API, process iteration and store result
-                // Streaming not yet supported for Responses API
+                // For streaming Responses API, leave execution_result in context
+                // for external streaming processor (serve_harmony_responses_stream)
                 if is_streaming {
-                    return Err(utils::internal_error_static(
-                        "Streaming not yet supported for Responses API",
-                    ));
+                    // Don't take execution_result - let the caller handle it
+                    return Ok(None);
                 }
 
+                // For non-streaming, process normally
+                let execution_result = ctx
+                    .state
+                    .response
+                    .execution_result
+                    .take()
+                    .ok_or_else(|| utils::internal_error_static("No execution result"))?;
+
+                let dispatch = ctx
+                    .state
+                    .dispatch
+                    .as_ref()
+                    .cloned()
+                    .ok_or_else(|| utils::internal_error_static("Dispatch metadata not set"))?;
+
                 let responses_request = ctx.responses_request_arc();
                 let iteration_result = self
                     .processor

sgl-router/src/routers/grpc/harmony/streaming.rs

@@ -16,20 +16,25 @@ use proto::{
 use serde_json::json;
 use tokio::sync::mpsc;
 use tokio_stream::{wrappers::UnboundedReceiverStream, StreamExt};
-use tracing::error;
+use tracing::{debug, error};
 
-use super::{types::HarmonyChannelDelta, HarmonyParserAdapter};
+use super::{
+    processor::ResponsesIterationResult, types::HarmonyChannelDelta, HarmonyParserAdapter,
+};
 use crate::{
     grpc_client::{proto, sglang_scheduler::AbortOnDropStream},
     protocols::{
         chat::{
             ChatCompletionRequest, ChatCompletionStreamResponse, ChatMessageDelta, ChatStreamChoice,
         },
-        common::{FunctionCallDelta, ToolCallDelta, Usage},
+        common::{FunctionCallDelta, ToolCall, ToolCallDelta, Usage},
+        responses::{ResponseStatus, ResponseUsage, ResponsesResponse, ResponsesUsage},
+    },
+    routers::grpc::{
+        context,
+        responses::streaming::{OutputItemType, ResponseStreamEventEmitter},
     },
-    routers::grpc::context,
 };
-
 /// Processor for streaming Harmony responses
 ///
 /// Returns an SSE stream that parses Harmony tokens incrementally and
@@ -526,6 +531,425 @@ impl HarmonyStreamingProcessor {
         Ok(())
     }
 
+    /// Common decode stream processing logic for both single and dual stream modes
+    ///
+    /// This helper function contains the shared logic for processing the decode stream,
+    /// parsing Harmony tokens, emitting SSE events, and tracking state.
+    async fn process_decode_stream_common(
+        mut decode_stream: AbortOnDropStream,
+        emitter: &mut ResponseStreamEventEmitter,
+        tx: &mpsc::UnboundedSender<Result<Bytes, io::Error>>,
+    ) -> Result<ResponsesIterationResult, String> {
+        // Initialize Harmony parser for this iteration
+        let mut parser =
+            HarmonyParserAdapter::new().map_err(|e| format!("Failed to create parser: {}", e))?;
+
+        // State tracking for channels
+        let mut has_analysis = false;
+        let mut accumulated_final_text = String::new();
+        let mut accumulated_tool_calls: Option<Vec<ToolCall>> = None;
+
+        // Track which items we've started
+        let mut reasoning_output_index: Option<usize> = None;
+        let mut message_output_index: Option<usize> = None;
+        let mut message_item_id: Option<String> = None;
+        let mut has_emitted_content_part_added = false;
+
+        // MCP tool call tracking (call_index -> (output_index, item_id))
+        let mut mcp_call_tracking: HashMap<usize, (usize, String)> = HashMap::new();
+
+        // Metadata from Complete message
+        let mut finish_reason = String::from("stop");
+        let mut matched_stop: Option<serde_json::Value> = None;
+
+        // Process stream
+        let mut chunk_count = 0;
+        while let Some(result) = decode_stream.next().await {
+            chunk_count += 1;
+            let response = result.map_err(|e| format!("Decode stream error: {}", e))?;
+
+            match response.response {
+                Some(Chunk(chunk)) => {
+                    // Parse chunk via Harmony parser
+                    let delta_result = parser
+                        .parse_chunk(&chunk.token_ids)
+                        .map_err(|e| format!("Parse error: {}", e))?;
+
+                    // Emit SSE events if there's a delta
+                    if let Some(delta) = delta_result {
+                        // Analysis channel → Reasoning item (wrapper events only, emitted once)
+                        if let Some(_analysis_text) = &delta.analysis_delta {
+                            if reasoning_output_index.is_none() {
+                                // Allocate reasoning item and emit wrapper events
+                                let (output_index, _item_id) =
+                                    emitter.allocate_output_index(OutputItemType::Reasoning);
+                                reasoning_output_index = Some(output_index);
+
+                                // Emit reasoning item (added + done in one call)
+                                // Note: reasoning_content will be provided at finalize
+                                emitter
+                                    .emit_reasoning_item(tx, None)
+                                    .map_err(|e| format!("Failed to emit reasoning item: {}", e))?;
+
+                                has_analysis = true;
+                            }
+                        }
+
+                        // Final channel → Message item (WITH text streaming)
+                        if let Some(final_delta) = &delta.final_delta {
+                            if !final_delta.is_empty() {
+                                // Allocate message item if needed
+                                if message_output_index.is_none() {
+                                    let (output_index, item_id) =
+                                        emitter.allocate_output_index(OutputItemType::Message);
+                                    message_output_index = Some(output_index);
+                                    message_item_id = Some(item_id.clone());
+
+                                    // Build message item structure
+                                    let item = json!({
+                                        "id": item_id,
+                                        "type": "message",
+                                        "role": "assistant",
+                                        "content": []
+                                    });
+
+                                    // Emit output_item.added
+                                    let event = emitter.emit_output_item_added(output_index, &item);
+                                    emitter.send_event_best_effort(&event, tx);
+                                }
+
+                                let output_index = message_output_index.unwrap();
+                                let item_id = message_item_id.as_ref().unwrap();
+                                let content_index = 0; // Single content part
+
+                                // Emit content_part.added before first delta
+                                if !has_emitted_content_part_added {
+                                    let event = emitter.emit_content_part_added(
+                                        output_index,
+                                        item_id,
+                                        content_index,
+                                    );
+                                    emitter.send_event_best_effort(&event, tx);
+                                    has_emitted_content_part_added = true;
+                                }
+
+                                // Emit text delta
+                                let event = emitter.emit_text_delta(
+                                    final_delta,
+                                    output_index,
+                                    item_id,
+                                    content_index,
+                                );
+                                emitter.send_event_best_effort(&event, tx);
+
+                                accumulated_final_text.push_str(final_delta);
+                            }
+                        }
+
+                        // Commentary channel → MCP tool call streaming
+                        if let Some(tc_delta) = &delta.commentary_delta {
+                            let call_index = tc_delta.index;
+
+                            // Check if this is a new tool call (has id and name)
+                            if tc_delta.id.is_some() {
+                                // NEW MCP CALL: Allocate output item and emit in_progress
+                                let (output_index, item_id) =
+                                    emitter.allocate_output_index(OutputItemType::McpCall);
+
+                                // Store tracking info
+                                mcp_call_tracking
+                                    .insert(call_index, (output_index, item_id.clone()));
+
+                                // Emit mcp_call.in_progress
+                                let event =
+                                    emitter.emit_mcp_call_in_progress(output_index, &item_id);
+                                emitter.send_event_best_effort(&event, tx);
+
+                                // If we have function name, emit initial mcp_call_arguments.delta
+                                if let Some(func) = &tc_delta.function {
+                                    if func.name.is_some() {
+                                        let event = emitter.emit_mcp_call_arguments_delta(
+                                            output_index,
+                                            &item_id,
+                                            "",
+                                        );
+                                        emitter.send_event_best_effort(&event, tx);
+                                    }
+                                }
+                            } else {
+                                // CONTINUING MCP CALL: Emit arguments delta
+                                if let Some((output_index, item_id)) =
+                                    mcp_call_tracking.get(&call_index)
+                                {
+                                    if let Some(args) = tc_delta
+                                        .function
+                                        .as_ref()
+                                        .and_then(|f| f.arguments.as_ref())
+                                        .filter(|a| !a.is_empty())
+                                    {
+                                        let event = emitter.emit_mcp_call_arguments_delta(
+                                            *output_index,
+                                            item_id,
+                                            args,
+                                        );
+                                        emitter.send_event_best_effort(&event, tx);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                Some(Complete(complete)) => {
+                    // Store final metadata
+                    finish_reason = complete.finish_reason.clone();
+                    matched_stop = complete.matched_stop.as_ref().map(|m| match m {
+                        MatchedTokenId(id) => {
+                            serde_json::json!(id)
+                        }
+                        MatchedStopStr(s) => {
+                            serde_json::json!(s)
+                        }
+                    });
+
+                    // Finalize parser and get complete output
+                    let final_output = parser
+                        .finalize(finish_reason.clone(), matched_stop.clone())
+                        .map_err(|e| format!("Finalize error: {}", e))?;
+
+                    // Store finalized tool calls
+                    accumulated_tool_calls = final_output.commentary.clone();
+
+                    // Complete all MCP tool calls if we have commentary
+                    if let Some(ref tool_calls) = accumulated_tool_calls {
+                        for (call_idx, tool_call) in tool_calls.iter().enumerate() {
+                            if let Some((output_index, item_id)) = mcp_call_tracking.get(&call_idx)
+                            {
+                                // Emit mcp_call_arguments.done with final arguments
+                                let args_str =
+                                    tool_call.function.arguments.as_deref().unwrap_or("");
+                                let event = emitter.emit_mcp_call_arguments_done(
+                                    *output_index,
+                                    item_id,
+                                    args_str,
+                                );
+                                emitter.send_event_best_effort(&event, tx);
+
+                                // Emit mcp_call.completed
+                                let event = emitter.emit_mcp_call_completed(*output_index, item_id);
+                                emitter.send_event_best_effort(&event, tx);
+
+                                // Mark output item as completed
+                                emitter.complete_output_item(*output_index);
+                            }
+                        }
+                    }
+
+                    // Close message item if we opened one
+                    if let Some(output_index) = message_output_index {
+                        let item_id = message_item_id.as_ref().unwrap();
+                        let content_index = 0;
+
+                        // Emit text_done
+                        let event = emitter.emit_text_done(output_index, item_id, content_index);
+                        emitter.send_event_best_effort(&event, tx);
+
+                        // Emit content_part.done
+                        let event =
+                            emitter.emit_content_part_done(output_index, item_id, content_index);
+                        emitter.send_event_best_effort(&event, tx);
+
+                        // Emit output_item.done
+                        let item = json!({
+                            "id": item_id,
+                            "type": "message",
+                            "role": "assistant",
+                            "content": [{
+                                "type": "text",
+                                "text": accumulated_final_text.clone()
+                            }]
+                        });
+                        let event = emitter.emit_output_item_done(output_index, &item);
+                        emitter.send_event_best_effort(&event, tx);
+
+                        emitter.complete_output_item(output_index);
+                    }
+                }
+                Some(proto::generate_response::Response::Error(err)) => {
+                    return Err(format!("Server error: {}", err.message));
+                }
+                None => {}
+            }
+        }
+
+        debug!(
+            "Stream loop ended. Total chunks received: {}, has_analysis: {}, tool_calls: {}, final_text_len: {}",
+            chunk_count,
+            has_analysis,
+            accumulated_tool_calls.as_ref().map(|tc| tc.len()).unwrap_or(0),
+            accumulated_final_text.len()
+        );
+
+        // Extract tool calls from completed messages or incomplete commentary
+        if chunk_count > 0 && accumulated_tool_calls.is_none() {
+            let messages = parser.get_messages();
+
+            // Try extracting from completed messages first
+            let (analysis_opt, commentary_opt, final_text_extracted) =
+                HarmonyParserAdapter::parse_messages(&messages);
+            accumulated_tool_calls = commentary_opt.clone();
+
+            // If no tool calls found, check for incomplete commentary in parser state
+            if accumulated_tool_calls.is_none() {
+                accumulated_tool_calls = parser.extract_incomplete_commentary();
+            }
+
+            debug!(
+                "Tool call extraction: completed_msgs={}, tool_calls={}, has_analysis={}, final_text_len={}",
+                messages.len(),
+                accumulated_tool_calls.as_ref().map(|tc| tc.len()).unwrap_or(0),
+                analysis_opt.is_some(),
+                final_text_extracted.len()
+            );
+
+            // Complete any pending MCP tool calls with data from completed messages
+            if let Some(ref tool_calls) = accumulated_tool_calls {
+                for (call_idx, tool_call) in tool_calls.iter().enumerate() {
+                    if let Some((output_index, item_id)) = mcp_call_tracking.get(&call_idx) {
+                        // Emit mcp_call_arguments.done with final arguments
+                        let args_str = tool_call.function.arguments.as_deref().unwrap_or("");
+                        let event =
+                            emitter.emit_mcp_call_arguments_done(*output_index, item_id, args_str);
+                        emitter.send_event_best_effort(&event, tx);
+
+                        // Emit mcp_call.completed
+                        let event = emitter.emit_mcp_call_completed(*output_index, item_id);
+                        emitter.send_event_best_effort(&event, tx);
+                    }
+                }
+            }
+        }
+
+        // Mark stream as completed successfully to prevent abort on drop
+        decode_stream.mark_completed();
+
+        // Return result based on whether tool calls were found
+        if let Some(tool_calls) = accumulated_tool_calls {
+            if !tool_calls.is_empty() {
+                let analysis_content = if has_analysis {
+                    // Get analysis from finalized parser output by calling finalize again
+                    // This is safe because finalize can be called multiple times
+                    let output = parser.finalize(finish_reason.clone(), matched_stop.clone())?;
+                    output.analysis
+                } else {
+                    None
+                };
+
+                return Ok(ResponsesIterationResult::ToolCallsFound {
+                    tool_calls,
+                    analysis: analysis_content,
+                    partial_text: accumulated_final_text,
+                });
+            }
+        }
+
+        // For streaming, we don't build the full ResponsesResponse here
+        // The caller will build it from the SSE events
+        // Return a placeholder Completed result (caller ignores these fields in streaming mode)
+        Ok(ResponsesIterationResult::Completed {
+            response: Box::new(ResponsesResponse {
+                id: String::new(),
+                object: "response".to_string(),
+                created_at: 0,
+                status: ResponseStatus::Completed,
+                error: None,
+                incomplete_details: None,
+                instructions: None,
+                max_output_tokens: None,
+                model: String::new(),
+                output: vec![],
+                parallel_tool_calls: true,
+                previous_response_id: None,
+                reasoning: None,
+                store: true,
+                temperature: None,
+                text: None,
+                tool_choice: "auto".to_string(),
+                tools: vec![],
+                top_p: None,
+                truncation: None,
+                user: None,
+                metadata: HashMap::new(),
+                usage: Some(ResponsesUsage::Modern(ResponseUsage {
+                    input_tokens: 0,
+                    output_tokens: 0,
+                    total_tokens: 0,
+                    input_tokens_details: None,
+                    output_tokens_details: None,
+                })),
+            }),
+            usage: Usage {
+                prompt_tokens: 0,
+                completion_tokens: 0,
+                total_tokens: 0,
+                completion_tokens_details: None,
+            },
+        })
+    }
+
+    /// Process streaming chunks for Responses API iteration
+    ///
+    /// Returns ResponsesIterationResult indicating whether tool calls were found
+    /// (requiring MCP loop continuation) or if the iteration is complete.
+    pub async fn process_responses_iteration_stream(
+        execution_result: context::ExecutionResult,
+        emitter: &mut ResponseStreamEventEmitter,
+        tx: &mpsc::UnboundedSender<Result<Bytes, io::Error>>,
+    ) -> Result<ResponsesIterationResult, String> {
+        match execution_result {
+            context::ExecutionResult::Single { stream } => {
+                debug!("Processing Responses API single stream mode");
+                Self::process_responses_single_stream(stream, emitter, tx).await
+            }
+            context::ExecutionResult::Dual { prefill, decode } => {
+                debug!("Processing Responses API dual stream mode");
+                Self::process_responses_dual_stream(prefill, *decode, emitter, tx).await
+            }
+        }
+    }
+
+    /// Process streaming chunks from a single stream (Responses API)
+    async fn process_responses_single_stream(
+        grpc_stream: AbortOnDropStream,
+        emitter: &mut ResponseStreamEventEmitter,
+        tx: &mpsc::UnboundedSender<Result<Bytes, io::Error>>,
+    ) -> Result<ResponsesIterationResult, String> {
+        // Delegate to common helper
+        Self::process_decode_stream_common(grpc_stream, emitter, tx).await
+    }
+
+    /// Process streaming chunks from dual streams (Responses API)
+    async fn process_responses_dual_stream(
+        mut prefill_stream: AbortOnDropStream,
+        decode_stream: AbortOnDropStream,
+        emitter: &mut ResponseStreamEventEmitter,
+        tx: &mpsc::UnboundedSender<Result<Bytes, io::Error>>,
+    ) -> Result<ResponsesIterationResult, String> {
+        // Phase 1: Process prefill stream (collect metadata, no output)
+        while let Some(result) = prefill_stream.next().await {
+            let _response = result.map_err(|e| format!("Prefill stream error: {}", e))?;
+            // No-op for prefill in Responses API (just metadata collection)
+        }
+
+        // Phase 2: Process decode stream using common helper
+        let result = Self::process_decode_stream_common(decode_stream, emitter, tx).await;
+
+        // Mark prefill stream as completed AFTER decode completes successfully
+        // This ensures that if client disconnects during decode, BOTH streams send abort
+        prefill_stream.mark_completed();
+
+        result
+    }
+
     /// Build SSE response from receiver
     fn build_sse_response(rx: mpsc::UnboundedReceiver<Result<Bytes, io::Error>>) -> Response {
         let stream = UnboundedReceiverStream::new(rx);

sgl-router/src/routers/grpc/pipeline.rs

@@ -454,4 +454,53 @@ impl RequestPipeline {
                 utils::internal_error_static("No ResponsesIterationResult produced by pipeline")
             })
     }
+
+    /// Execute Harmony Responses pipeline iteration with streaming support
+    ///
+    /// This version executes the pipeline up to the dispatch stage and returns
+    /// the raw ExecutionResult (with stream) for token-level streaming processing.
+    pub async fn execute_harmony_responses_streaming(
+        &self,
+        request: &crate::protocols::responses::ResponsesRequest,
+        harmony_ctx: &harmony::responses::HarmonyResponsesContext,
+    ) -> Result<ExecutionResult, Response> {
+        // Create RequestContext for this Responses request
+        let mut ctx = RequestContext::for_responses(
+            Arc::new(request.clone()),
+            None,
+            None,
+            harmony_ctx.components.clone(),
+        );
+
+        // Execute pipeline stages up to dispatch (which creates the stream)
+        for (idx, stage) in self.stages.iter().enumerate() {
+            match stage.execute(&mut ctx).await {
+                Ok(Some(response)) => {
+                    error!(
+                        "Stage {} ({}) returned unexpected response during streaming Responses",
+                        idx + 1,
+                        stage.name()
+                    );
+                    return Err(response);
+                }
+                Ok(None) => continue,
+                Err(response) => {
+                    error!(
+                        "Stage {} ({}) failed with status {}",
+                        idx + 1,
+                        stage.name(),
+                        response.status()
+                    );
+                    return Err(response);
+                }
+            }
+        }
+
+        // Extract execution_result (the raw stream from workers)
+        ctx.state
+            .response
+            .execution_result
+            .take()
+            .ok_or_else(|| utils::internal_error_static("No ExecutionResult produced by pipeline"))
+    }
 }

sgl-router/src/routers/grpc/responses/streaming.rs

@@ -9,7 +9,7 @@ use uuid::Uuid;
 
 use crate::protocols::chat::ChatCompletionStreamResponse;
 
-pub(super) enum OutputItemType {
+pub enum OutputItemType {
     Message,
     McpListTools,
     McpCall,
@@ -53,9 +53,9 @@ struct OutputItemState {
 /// - response.mcp_call_arguments.done
 /// - response.mcp_call.completed
 /// - response.mcp_call.failed
-pub(super) struct ResponseStreamEventEmitter {
+pub struct ResponseStreamEventEmitter {
     sequence_number: u64,
-    response_id: String,
+    pub response_id: String,
     model: String,
     created_at: u64,
     message_id: String,
@@ -74,7 +74,7 @@ pub(super) struct ResponseStreamEventEmitter {
 }
 
 impl ResponseStreamEventEmitter {
-    pub(super) fn new(response_id: String, model: String, created_at: u64) -> Self {
+    pub fn new(response_id: String, model: String, created_at: u64) -> Self {
         let message_id = format!("msg_{}", Uuid::new_v4());
 
         Self {
@@ -102,7 +102,7 @@ impl ResponseStreamEventEmitter {
         seq
     }
 
-    pub(super) fn emit_created(&mut self) -> serde_json::Value {
+    pub fn emit_created(&mut self) -> serde_json::Value {
         self.has_emitted_created = true;
         json!({
             "type": "response.created",
@@ -118,7 +118,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_in_progress(&mut self) -> serde_json::Value {
+    pub fn emit_in_progress(&mut self) -> serde_json::Value {
         self.has_emitted_in_progress = true;
         json!({
             "type": "response.in_progress",
@@ -131,7 +131,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_content_part_added(
+    pub fn emit_content_part_added(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -151,7 +151,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_text_delta(
+    pub fn emit_text_delta(
         &mut self,
         delta: &str,
         output_index: usize,
@@ -169,7 +169,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_text_done(
+    pub fn emit_text_done(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -185,7 +185,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_content_part_done(
+    pub fn emit_content_part_done(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -204,10 +204,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_completed(
-        &mut self,
-        usage: Option<&serde_json::Value>,
-    ) -> serde_json::Value {
+    pub fn emit_completed(&mut self, usage: Option<&serde_json::Value>) -> serde_json::Value {
         let mut response = json!({
             "type": "response.completed",
             "sequence_number": self.next_sequence(),
@@ -240,10 +237,7 @@ impl ResponseStreamEventEmitter {
     // MCP Event Emission Methods
     // ========================================================================
 
-    pub(super) fn emit_mcp_list_tools_in_progress(
-        &mut self,
-        output_index: usize,
-    ) -> serde_json::Value {
+    pub fn emit_mcp_list_tools_in_progress(&mut self, output_index: usize) -> serde_json::Value {
         json!({
             "type": "response.mcp_list_tools.in_progress",
             "sequence_number": self.next_sequence(),
@@ -251,7 +245,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_mcp_list_tools_completed(
+    pub fn emit_mcp_list_tools_completed(
         &mut self,
         output_index: usize,
         tools: &[crate::mcp::Tool],
@@ -275,7 +269,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_mcp_call_in_progress(
+    pub fn emit_mcp_call_in_progress(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -288,7 +282,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_mcp_call_arguments_delta(
+    pub fn emit_mcp_call_arguments_delta(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -309,7 +303,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_mcp_call_arguments_done(
+    pub fn emit_mcp_call_arguments_done(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -324,7 +318,7 @@ impl ResponseStreamEventEmitter {
         })
     }
 
-    pub(super) fn emit_mcp_call_completed(
+    pub fn emit_mcp_call_completed(
         &mut self,
         output_index: usize,
         item_id: &str,
@@ -357,7 +351,7 @@ impl ResponseStreamEventEmitter {
     // ========================================================================
 
     /// Emit response.output_item.added event
-    pub(super) fn emit_output_item_added(
+    pub fn emit_output_item_added(
         &mut self,
         output_index: usize,
         item: &serde_json::Value,
@@ -371,7 +365,7 @@ impl ResponseStreamEventEmitter {
     }
 
     /// Emit response.output_item.done event
-    pub(super) fn emit_output_item_done(
+    pub fn emit_output_item_done(
         &mut self,
         output_index: usize,
         item: &serde_json::Value,
@@ -390,7 +384,7 @@ impl ResponseStreamEventEmitter {
     }
 
     /// Allocate next output index and track item
-    pub(super) fn allocate_output_index(&mut self, item_type: OutputItemType) -> (usize, String) {
+    pub fn allocate_output_index(&mut self, item_type: OutputItemType) -> (usize, String) {
         let index = self.next_output_index;
         self.next_output_index += 1;
 
@@ -412,7 +406,7 @@ impl ResponseStreamEventEmitter {
     }
 
     /// Mark output item as completed
-    pub(super) fn complete_output_item(&mut self, output_index: usize) {
+    pub fn complete_output_item(&mut self, output_index: usize) {
         if let Some(item) = self
             .output_items
             .iter_mut()
@@ -426,7 +420,7 @@ impl ResponseStreamEventEmitter {
     ///
     /// Reasoning items in OpenAI format are simple placeholders emitted between tool iterations.
     /// They don't have streaming content - just wrapper events with empty/null content.
-    pub(super) fn emit_reasoning_item(
+    pub fn emit_reasoning_item(
         &mut self,
         tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
         reasoning_content: Option<String>,
@@ -550,7 +544,7 @@ impl ResponseStreamEventEmitter {
         Ok(())
     }
 
-    pub(super) fn send_event(
+    pub fn send_event(
         &self,
         event: &serde_json::Value,
         tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
@@ -558,13 +552,38 @@ impl ResponseStreamEventEmitter {
         let event_json = serde_json::to_string(event)
             .map_err(|e| format!("Failed to serialize event: {}", e))?;
 
-        if tx
-            .send(Ok(Bytes::from(format!("data: {}\n\n", event_json))))
-            .is_err()
-        {
+        // Extract event type from the JSON for SSE event field
+        let event_type = event
+            .get("type")
+            .and_then(|v| v.as_str())
+            .unwrap_or("message");
+
+        // Format as SSE with event: field
+        let sse_message = format!("event: {}\ndata: {}\n\n", event_type, event_json);
+
+        if tx.send(Ok(Bytes::from(sse_message))).is_err() {
             return Err("Client disconnected".to_string());
         }
 
         Ok(())
     }
+
+    /// Send event and log any errors (typically client disconnect)
+    ///
+    /// This is a convenience method for streaming scenarios where client
+    /// disconnection is expected and should be logged but not fail the operation.
+    /// Returns true if sent successfully, false if client disconnected.
+    pub fn send_event_best_effort(
+        &self,
+        event: &serde_json::Value,
+        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
+    ) -> bool {
+        match self.send_event(event, tx) {
+            Ok(()) => true,
+            Err(e) => {
+                tracing::debug!("Failed to send event (likely client disconnect): {}", e);
+                false
+            }
+        }
+    }
 }

sgl-router/src/routers/grpc/router.rs

@@ -13,7 +13,10 @@ use tracing::debug;
 
 use super::{
     context::SharedComponents,
-    harmony::{serve_harmony_responses, HarmonyDetector, HarmonyResponsesContext},
+    harmony::{
+        serve_harmony_responses, serve_harmony_responses_stream, HarmonyDetector,
+        HarmonyResponsesContext,
+    },
     pipeline::RequestPipeline,
     responses,
 };
@@ -192,8 +195,8 @@ impl GrpcRouter {
         model_id: Option<&str>,
     ) -> Response {
         debug!(
-            "Processing Harmony responses request for model: {:?}",
-            model_id
+            "Processing Harmony responses request for model: {:?}, streaming: {:?}",
+            model_id, body.stream
         );
 
         // Create HarmonyResponsesContext from existing responses context
@@ -204,10 +207,15 @@ impl GrpcRouter {
             self.harmony_responses_context.response_storage.clone(),
         );
 
-        // Use serve_harmony_responses for multi-turn MCP tool orchestration
-        match serve_harmony_responses(&harmony_ctx, body.clone()).await {
-            Ok(response) => axum::Json(response).into_response(),
-            Err(error_response) => error_response,
+        // Check if streaming is requested
+        if body.stream.unwrap_or(false) {
+            serve_harmony_responses_stream(&harmony_ctx, body.clone()).await
+        } else {
+            // Use non-streaming version for standard JSON responses
+            match serve_harmony_responses(&harmony_ctx, body.clone()).await {
+                Ok(response) => axum::Json(response).into_response(),
+                Err(error_response) => error_response,
+            }
         }
     }
 }