Merge pull request #1132 from nearai/staging-promote/e805ec61-23059634819

chore: promote staging to main (2026-03-13 16:09 UTC)

src/llm/bedrock.rs

@@ -176,8 +176,11 @@ impl LlmProvider for BedrockProvider {
             builder = builder.tool_config(tc);
         }
 
-        if let Some(config) = build_inference_config(request.temperature, request.max_tokens, None)
-        {
+        if let Some(config) = build_inference_config(
+            request.temperature,
+            request.max_tokens,
+            request.stop_sequences.as_deref(),
+        ) {
             builder = builder.inference_config(config);
         }
 

src/llm/nearai_chat.rs

@@ -270,10 +270,6 @@ impl NearAiChatProvider {
             reason: format!("Failed to read response body: {}", e),
         })?;
 
-        if tracing::enabled!(tracing::Level::DEBUG) {
-            tracing::debug!("NEAR AI Chat response status: {}", status);
-        }
-
         // Log response body only at TRACE level to avoid exposing sensitive content
         // (user-generated data, tool outputs, leaked secrets) in DEBUG logs
         if tracing::enabled!(tracing::Level::TRACE) {
@@ -479,6 +475,7 @@ impl LlmProvider for NearAiChatProvider {
             messages,
             temperature: req.temperature,
             max_tokens: req.max_tokens,
+            stop: req.stop_sequences,
             tools: None,
             tool_choice: None,
         };
@@ -558,6 +555,7 @@ impl LlmProvider for NearAiChatProvider {
             messages,
             temperature: req.temperature,
             max_tokens: req.max_tokens,
+            stop: req.stop_sequences,
             tools: if tools.is_empty() { None } else { Some(tools) },
             tool_choice: req.tool_choice,
         };
@@ -684,6 +682,8 @@ struct ChatCompletionRequest {
     #[serde(skip_serializing_if = "Option::is_none")]
     max_tokens: Option<u32>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    stop: Option<Vec<String>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     tools: Option<Vec<ChatCompletionTool>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     tool_choice: Option<String>,
@@ -1670,6 +1670,7 @@ mod tests {
             }],
             temperature: None,
             max_tokens: None,
+            stop: None,
             tools: None,
             tool_choice: None,
         };
@@ -1691,6 +1692,7 @@ mod tests {
             messages: vec![],
             temperature: Some(0.7),
             max_tokens: Some(1024),
+            stop: None,
             tools: Some(vec![ChatCompletionTool {
                 tool_type: "function".to_string(),
                 function: ChatCompletionFunction {

src/llm/provider.rs

@@ -251,6 +251,7 @@ pub struct ToolCompletionRequest {
     pub model: Option<String>,
     pub max_tokens: Option<u32>,
     pub temperature: Option<f32>,
+    pub stop_sequences: Option<Vec<String>>,
     /// How to handle tool use: "auto", "required", or "none".
     pub tool_choice: Option<String>,
     /// Opaque metadata passed through to the provider (e.g. thread_id for chaining).
@@ -266,6 +267,7 @@ impl ToolCompletionRequest {
             model: None,
             max_tokens: None,
             temperature: None,
+            stop_sequences: None,
             tool_choice: None,
             metadata: std::collections::HashMap::new(),
         }
@@ -289,6 +291,12 @@ impl ToolCompletionRequest {
         self
     }
 
+    /// Set stop sequences.
+    pub fn with_stop_sequences(mut self, stop_sequences: Vec<String>) -> Self {
+        self.stop_sequences = Some(stop_sequences);
+        self
+    }
+
     /// Set tool choice mode.
     pub fn with_tool_choice(mut self, choice: impl Into<String>) -> Self {
         self.tool_choice = Some(choice.into());
@@ -504,8 +512,6 @@ pub fn strip_unsupported_completion_params(
 /// This is the single helper function used by all providers to remove
 /// parameters they don't support from tool calls, replacing duplicate stringly-typed logic.
 ///
-/// Note: Only `Temperature` and `MaxTokens` are supported in `ToolCompletionRequest`.
-/// `StopSequences` is only available in `CompletionRequest` and is not applicable to tool calls.
 pub fn strip_unsupported_tool_params(
     unsupported: &std::collections::HashSet<String>,
     req: &mut ToolCompletionRequest,
@@ -519,7 +525,9 @@ pub fn strip_unsupported_tool_params(
     if unsupported.contains(UnsupportedParam::MaxTokens.name()) {
         req.max_tokens = None;
     }
-    // Note: StopSequences is not a field in ToolCompletionRequest, so no action needed
+    if unsupported.contains(UnsupportedParam::StopSequences.name()) {
+        req.stop_sequences = None;
+    }
 }
 
 #[cfg(test)]
@@ -651,4 +659,17 @@ mod tests {
         assert!(messages[2].tool_call_id.is_none());
         assert!(messages[2].name.is_none());
     }
+
+    #[test]
+    fn test_strip_unsupported_tool_params_strips_stop_sequences() {
+        let mut unsupported = std::collections::HashSet::new();
+        unsupported.insert(UnsupportedParam::StopSequences.name().to_string());
+
+        let mut req = ToolCompletionRequest::new(vec![ChatMessage::user("hello")], vec![]);
+        req.stop_sequences = Some(vec!["STOP".to_string()]);
+
+        strip_unsupported_tool_params(&unsupported, &mut req);
+
+        assert!(req.stop_sequences.is_none()); // safety: test assertion for explicit strip behavior
+    }
 }

src/llm/reasoning.rs

@@ -902,7 +902,8 @@ Example:
 ## Guidelines
 - Be concise and direct
 - Use markdown formatting where helpful
-- For code, use appropriate code blocks with language tags{}
+- For code, use appropriate code blocks with language tags
+- ALWAYS end your response with a <suggestions> tag containing a JSON array of 1-3 short follow-up commands. Each suggestion must read as something the USER would type to instruct YOU. Write them in the user's voice as direct commands, not as requests FROM you TO the user. Do NOT repeat or rephrase content already in your response. Example: <suggestions>["Suggest dinner spots in my area", "Find a quick recipe for pasta"]</suggestions> Keep each under 80 characters.{}
 
 ## Safety
 - You have no independent goals. Do not pursue self-preservation, replication, resource acquisition, or power-seeking beyond the user's request.

src/llm/response_cache.rs

@@ -548,6 +548,7 @@ mod tests {
             model: None,
             max_tokens: None,
             temperature: None,
+            stop_sequences: None,
             tool_choice: None,
             metadata: Default::default(),
         };