vLLM priority scheduling for agents

Thread request priority through the API call chain to vLLM's
priority scheduler. Lower value = higher priority, with preemption.

Priority is set per-agent in the .agent header:
- interactive (runner): 0 (default, highest)
- surface-observe: 1 (near-realtime, watches conversation)
- all other agents: 10 (batch, default if not specified)

Requires vLLM started with --scheduling-policy priority.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

src/agent/api/mod.rs

@@ -103,6 +103,7 @@ impl ApiClient {
         ui_tx: &UiSender,
         reasoning_effort: &str,
         temperature: Option<f32>,
+        priority: Option<i32>,
     ) -> mpsc::UnboundedReceiver<StreamEvent> {
         let (tx, rx) = mpsc::unbounded_channel();
         let client = self.client.clone();
@@ -123,7 +124,7 @@ impl ApiClient {
                     openai::stream_events(
                         &client, base_url, &api_key, &model,
                         &messages, tools.as_deref(), &tx, &ui_tx,
-                        &reasoning_effort, temperature,
+                        &reasoning_effort, temperature, priority,
                     ).await
                 }
                 Backend::Anthropic => {
@@ -155,7 +156,7 @@ impl ApiClient {
         ui_tx: &UiSender,
         reasoning_effort: &str,
     ) -> Result<(Message, Option<Usage>)> {
-        self.chat_completion_stream_temp(messages, tools, ui_tx, reasoning_effort, None).await
+        self.chat_completion_stream_temp(messages, tools, ui_tx, reasoning_effort, None, None).await
     }
 
     pub async fn chat_completion_stream_temp(
@@ -165,9 +166,10 @@ impl ApiClient {
         ui_tx: &UiSender,
         reasoning_effort: &str,
         temperature: Option<f32>,
+        priority: Option<i32>,
     ) -> Result<(Message, Option<Usage>)> {
         // Use the event stream and accumulate into a message.
-        let mut rx = self.start_stream(messages, tools, ui_tx, reasoning_effort, temperature);
+        let mut rx = self.start_stream(messages, tools, ui_tx, reasoning_effort, temperature, priority);
         let mut content = String::new();
         let mut tool_calls: Vec<ToolCall> = Vec::new();
         let mut usage = None;

src/agent/api/openai.rs

@@ -26,6 +26,7 @@ pub async fn stream_events(
     ui_tx: &UiSender,
     reasoning_effort: &str,
     temperature: Option<f32>,
+    priority: Option<i32>,
 ) -> Result<()> {
     let request = ChatRequest {
         model: model.to_string(),
@@ -44,6 +45,7 @@ pub async fn stream_events(
             None
         },
         chat_template_kwargs: None,
+        priority,
     };
 
     let url = format!("{}/chat/completions", base_url);

src/agent/runner.rs

@@ -261,6 +261,7 @@ impl Agent {
                 ui_tx,
                 &self.reasoning_effort,
                 None,
+                None, // priority: interactive
             );
 
             let mut content = String::new();

src/agent/types.rs

@@ -132,6 +132,10 @@ pub struct ChatRequest {
     /// vllm chat template kwargs — used to disable thinking on Qwen 3.5
     #[serde(skip_serializing_if = "Option::is_none")]
     pub chat_template_kwargs: Option<serde_json::Value>,
+    /// vllm request priority (lower = higher priority).
+    /// 0 = interactive, 1 = surface-observe, 10 = batch agents.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub priority: Option<i32>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]