@@ -56,7 +56,7 @@ private async Task<RoleDialogModel> InnerGetChatCompletions(Agent agent, List<Ro
         }
         else if (reason == ChatFinishReason.Length)
         {
-            _logger.LogWarning($"Action: {nameof(InnerGetChatCompletions)}, Reason: {reason}, Agent: {agent.Name}, MaxOutputTokens: {options.MaxOutputTokenCount}, Content:{text}");
+            _logger.LogWarning($"Action: {nameof(InnerGetChatCompletions)}, Reason: {reason}, Agent: {agent.Name}, MaxOutputTokens: {options.MaxOutputTokenCount}, Content:{DescribeContent(text)}, ToolCalls: {DescribeToolCalls(ToLlmToolCalls(value.ToolCalls))}");
 
             responseMessage = new RoleDialogModel(AgentRole.Assistant, $"AI response exceeded max output length")
             {
@@ -176,7 +176,7 @@ private async Task<bool> InnerGetChatCompletionsAsync(Agent agent,
         }
         else if (reason == ChatFinishReason.Length)
         {
-            _logger.LogWarning($"Action: {nameof(InnerGetChatCompletionsAsync)}, Reason: {reason}, Agent: {agent.Name}, MaxOutputTokens: {options.MaxOutputTokenCount}, Content:{text}");
+            _logger.LogWarning($"Action: {nameof(InnerGetChatCompletionsAsync)}, Reason: {reason}, Agent: {agent.Name}, MaxOutputTokens: {options.MaxOutputTokenCount}, Content:{DescribeContent(text)}, ToolCalls: {DescribeToolCalls(ToLlmToolCalls(value.ToolCalls))}");
 
             msg = new RoleDialogModel(AgentRole.Assistant, $"AI response exceeded max output length")
             {
@@ -321,6 +321,13 @@ private async Task<RoleDialogModel> InnerGetChatCompletionsStreamingAsync(Agent
                 }
                 else if (choice.FinishReason.HasValue)
                 {
+                    if (choice.FinishReason == ChatFinishReason.Length)
+                    {
+                        // What the model wrote before the limit is replaced by the notice below, so this is the only record of it
+                        _logger.LogWarning($"Action: {nameof(InnerGetChatCompletionsStreamingAsync)}, Reason: {choice.FinishReason}, Agent: {agent.Name}, " +
+                            $"MaxOutputTokens: {options.MaxOutputTokenCount}, Content:{DescribeContent(textStream.GetText())}, ToolCalls: {DescribeToolCalls(ReconstructToolCalls(toolCalls))}");
+                    }
+
                     var text = choice.FinishReason == ChatFinishReason.Length ? "Model reached the maximum number of tokens allowed."
                         : choice.FinishReason == ChatFinishReason.ContentFilter ? "Content is omitted due to content filter rule."
                         : choice.FinishReason.Value.ToString();
@@ -759,6 +766,20 @@ private static List<LlmToolCall> ToLlmToolCalls(IEnumerable<ChatToolCall>? toolC
             .Select(x => new LlmToolCall(x.Id, x.FunctionName, x.FunctionArguments?.ToString()))
             .ToList();
 
+    /// <summary>
+    /// The tool calls of a reply cut off by the output limit: each name and how long its arguments got. The arguments
+    /// themselves are left out, since they can hold the user's personal data.
+    /// </summary>
+    internal static string DescribeToolCalls(List<LlmToolCall> calls)
+        => string.Join(", ", calls.Select(x => $"{x.FunctionName}({x.FunctionArgs?.Length ?? 0} chars)"));
+
+    /// <summary>
+    /// What a reply cut off by the output limit said: its length and only its start, enough to tell what the model was
+    /// writing without logging the whole reply.
+    /// </summary>
+    internal static string DescribeContent(string? text)
+        => string.IsNullOrEmpty(text) ? "(none)" : $"{text.Length} chars: {(text.Length > 200 ? text[..200] + "..." : text)}";
+
     /// <summary>
     /// Rebuilds the tool calls a streaming reply asked for.
     /// </summary>
 
@@ -93,6 +93,23 @@ await HookEmitter.Emit<ITwilioCallStatusHook>(_services,
             {
                 response = twilio.ReturnBidirectionalMediaStreamsInstructions(instruction, agent);
             }
+            // A call that arrives with an intent already has its first message queued for the Gather flow
+            else if (string.IsNullOrWhiteSpace(request.Intent) && _settings.ConversationRelay.IsEnabledFor(request.AgentId))
+            {
+                // ConversationRelay speaks text only: a greeting a session hook wrote (e.g. with the caller's name) is
+                // said as it is, otherwise the prompts the session hooks queued become the greeting
+                var greeting = !string.IsNullOrWhiteSpace(instruction.Text)
+                    ? instruction.Text
+                    : string.Join(" ", instruction.SpeechPaths
+                        .Select(twilio.GetPromptText)
+                        .Where(text => !string.IsNullOrWhiteSpace(text)));
+                if (greeting.Length == 0 && instruction.SpeechPaths.Count > 0)
+                {
+                    _logger.LogWarning($"No prompt text for the welcome speech of agent {request.AgentId}; " +
+                        "starting ConversationRelay without a greeting.");
+                }
+                response = twilio.ReturnConversationRelay(request.AgentId, request.ConversationId, greeting);
+            }
             else
             {
                 if (string.IsNullOrWhiteSpace(request.Intent))
 
@@ -21,4 +21,10 @@ public class AgentTurnCallbacks
     /// Called when the model starts writing a new response, e.g. after a tool call
     /// </summary>
     public Func<Task>? OnStreamStarted { get; init; }
+
+    /// <summary>
+    /// Called when a response the model wrote ends in a tool call: what it said before the call is complete,
+    /// and the tool may take a while before anything else is said
+    /// </summary>
+    public Func<Task>? OnStreamEndedWithToolCall { get; init; }
 }