First cut of additional AI_DEBUG

app/models/assistant.rb

@@ -55,6 +55,37 @@ class Assistant
     responder.on(:response) do |data|
       update_thinking("Analyzing your data...")
 
+      # Persist the provider's response identifier on the assistant message so
+      # future renders reflect the exact metadata used for this conversation
+      assistant_message.update!(provider_id: data[:id]) if data[:id].present?
+
+      # Persist the endpoint used for this provider (if applicable)
+      if assistant_message.endpoint.blank? && llm_provider.respond_to?(:endpoint_base)
+        assistant_message.update!(endpoint: llm_provider.endpoint_base)
+      end
+
+      # Persist usage metrics and estimated cost when provided by the LLM provider
+      if data[:usage].present?
+        usage = data[:usage]
+
+        prompt_tokens = usage["prompt_tokens"] || usage["input_tokens"] || 0
+        completion_tokens = usage["completion_tokens"] || usage["output_tokens"] || 0
+        total_tokens = usage["total_tokens"] || (prompt_tokens + completion_tokens)
+
+        estimated_cost = LlmUsage.calculate_cost(
+          model: message.ai_model,
+          prompt_tokens: prompt_tokens,
+          completion_tokens: completion_tokens
+        )
+
+        assistant_message.update!(
+          prompt_tokens: prompt_tokens,
+          completion_tokens: completion_tokens,
+          total_tokens: total_tokens,
+          estimated_cost: estimated_cost
+        )
+      end
+
       if data[:function_tool_calls].present?
         assistant_message.tool_calls = data[:function_tool_calls]
         latest_response_id = data[:id]

app/models/assistant/responder.rb

@@ -24,9 +24,9 @@ class Assistant::Responder
         response_handled = true
 
         if response.function_requests.any?
-          handle_follow_up_response(response)
+          handle_follow_up_response(response, usage: chunk.usage)
         else
-          emit(:response, { id: response.id })
+          emit(:response, { id: response.id, usage: chunk.usage })
         end
       end
     end
@@ -46,14 +46,14 @@ class Assistant::Responder
   private
     attr_reader :message, :instructions, :function_tool_caller, :llm
 
-    def handle_follow_up_response(response)
+    def handle_follow_up_response(response, usage: nil)
       streamer = proc do |chunk|
         case chunk.type
         when "output_text"
           emit(:output_text, chunk.data)
         when "response"
           # We do not currently support function executions for a follow-up response (avoid recursive LLM calls that could lead to high spend)
-          emit(:response, { id: chunk.data.id })
+          emit(:response, { id: chunk.data.id, usage: chunk.usage })
         end
       end
 
@@ -61,6 +61,7 @@ class Assistant::Responder
 
       emit(:response, {
         id: response.id,
+        usage: usage,
         function_tool_calls: function_tool_calls
       })
 
@@ -70,7 +71,7 @@ class Assistant::Responder
         function_results: function_tool_calls.map(&:to_result),
         previous_response_id: response.id
       )
-    end
+      end
 
     def get_llm_response(streamer:, function_results: [], previous_response_id: nil)
       response = llm.chat_response(

app/models/provider/openai.rb

@@ -39,6 +39,12 @@ class Provider::Openai < Provider
     custom_provider? ? "Custom OpenAI-compatible (#{@uri_base})" : "OpenAI"
   end
 
+  # Returns the HTTP base endpoint used by this provider, or nil when using
+  # the default OpenAI endpoint configured by the SDK.
+  def endpoint_base
+    @uri_base.presence
+  end
+
   def supported_models_description
     if custom_provider?
       @default_model.present? ? "configured model: #{@default_model}" : "any model"