fix(loop): re-enable loop intervention system messages

**Problem:**
Claude and other models were looping endlessly when stuck (calling same tool 25+ times).
Loop detection was working but intervention message injection was DISABLED due to previous
GitHub Copilot API 400 errors.

**Root Cause:**
Comment in code: "system message injection removed (was causing GitHub Copilot 400 errors)"
The 400 errors were caused by bugs that were FIXED in previous commits:
- Delta mode fixes (d1c3ea7)
- User role requirement (5fff807)
- Trailing whitespace (822c12c)

**Solution:**
Re-enabled system-reminder intervention message injection in both:
1. Basic loop detection path (AgentOrchestrator.swift line ~2318)
2. Streaming loop detection path (AgentOrchestrator.swift line ~3668)

When loop detected (3+ consecutive calls to same tool):
- Creates intervention message explaining the loop
- Wraps in <system-reminder> tags (Claude-compatible format)
- Appends to internalMessages for next API request
- Model receives warning and changes behavior

**Testing:**
✅ Build: PASS
✅ Claude: Stops looping after intervention (was 25+ calls, now stops at 3)
✅ GPT-4.1: Works correctly, no regressions
✅ No API errors (400s no longer occur due to previous fixes)
✅ Loop intervention message appears in API request
✅ Model adapts behavior after receiving intervention

Author: fewtarius

Sources/APIFramework/AgentOrchestrator.swift

@@ -2320,8 +2320,22 @@ public class AgentOrchestrator: ObservableObject, IterationController {
                     if count >= limit {
                         logger.warning("LOOP_DETECTION: Tool '\(toolName)' called \(count) times consecutively! Loop detected.")
 
+                        /// Create intervention message to break the loop
+                        let interventionMessage = """
+                        LOOP DETECTED: You've called the '\(toolName)' tool \(count) times in a row without progress.
+                        
+                        This indicates you may be stuck. Please:
+                        1. Review the results you've already received
+                        2. Try a different approach or tool
+                        3. If the task cannot be completed, explain why to the user
+                        
+                        Do NOT call '\(toolName)' again immediately - choose a different action.
+                        """
+                        
+                        context.internalMessages.append(createSystemReminder(content: interventionMessage, model: model))
+                        
                         context.consecutiveToolCounts[toolName] = 0
-                        logger.debug("LOOP_DETECTION: Detected '\(toolName)' loop (system message injection removed)")
+                        logger.debug("LOOP_DETECTION: Injected intervention message for '\(toolName)' loop")
                     }
                 }
 
@@ -3658,8 +3672,22 @@ public class AgentOrchestrator: ObservableObject, IterationController {
                             if count >= limit {
                                 logger.warning("LOOP_DETECTION_STREAMING: Tool '\(toolName)' called \(count) times consecutively! Loop detected.")
 
+                                /// Create intervention message to break the loop
+                                let interventionMessage = """
+                                LOOP DETECTED: You've called the '\(toolName)' tool \(count) times in a row without progress.
+                                
+                                This indicates you may be stuck. Please:
+                                1. Review the results you've already received
+                                2. Try a different approach or tool
+                                3. If the task cannot be completed, explain why to the user
+                                
+                                Do NOT call '\(toolName)' again immediately - choose a different action.
+                                """
+                                
+                                context.internalMessages.append(createSystemReminder(content: interventionMessage, model: model))
+                                
                                 context.consecutiveToolCounts[toolName] = 0
-                                logger.debug("LOOP_DETECTION_STREAMING: Injected '\(toolName)' loop intervention message")
+                                logger.debug("LOOP_DETECTION_STREAMING: Injected intervention message for '\(toolName)' loop")
                             }
                         }
 
@@ -5148,6 +5176,13 @@ public class AgentOrchestrator: ObservableObject, IterationController {
             for (index, historyMessage) in conversationMessages.enumerated() {
                 var cleanContent = historyMessage.content
 
+                /// DIAGNOSTIC: Track isToolMessage flag at conversion
+                let hasPreview = cleanContent.contains("[TOOL_RESULT_PREVIEW]") || cleanContent.contains("[TOOL_RESULT_STORED]")
+                if hasPreview {
+                    let contentPrefix = cleanContent.prefix(60).replacingOccurrences(of: "\n", with: " ")
+                    logger.debug("CONVERT_TOOL_MSG: isToolMessage=\(historyMessage.isToolMessage), type=\(historyMessage.type), contentPrefix=[\(contentPrefix)]")
+                }
+
                 // Handle tool messages with proper role and toolCallId
                 if historyMessage.isToolMessage {
                     // Tool messages need role="tool" and toolCallId for proper API formatting

Sources/ConversationEngine/ConversationMessageBus.swift

@@ -567,6 +567,10 @@ public class ConversationMessageBus: ObservableObject {
             isToolMessage: true
         )
 
+        /// DIAGNOSTIC: Track isToolMessage flag at creation
+        let contentPrefix = message.content.prefix(60).replacingOccurrences(of: "\n", with: " ")
+        logger.debug("TOOL_MESSAGE_CREATED: id=\(message.id), isToolMessage=\(message.isToolMessage), contentPrefix=[\(contentPrefix)]")
+
         appendMessage(message)
         scheduleSave()