feat(agent): add kill-switch for final-answer Markdown normalization

Gate MarkdownNormalizer behind mate.agent.markdown-normalize-enabled (default
true) so operators can disable the rewrite verbatim if a normalization edge
case ever mangles a legitimate answer.

mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java

@@ -88,6 +88,11 @@ public class AgentGraphBuilder {
     @org.springframework.beans.factory.annotation.Value(
             "${mateclaw.skill.disclosure.load-skill-tool.enabled:true}")
     private boolean loadSkillToolEnabled;
+
+    /** Escape hatch: when false, the final answer is sent verbatim without Markdown normalization. */
+    @org.springframework.beans.factory.annotation.Value(
+            "${mate.agent.markdown-normalize-enabled:true}")
+    private boolean markdownNormalizeEnabled;
     private final ConversationService conversationService;
     private final ModelConfigService modelConfigService;
     private final ModelProviderService modelProviderService;
@@ -809,7 +814,7 @@ public class AgentGraphBuilder {
             SummarizingNode summarizingNode = new SummarizingNode(chatModel, streamingHelper, streamTracker);
             LimitExceededNode limitExceededNode = new LimitExceededNode(
                     chatModel, observationProcessor, streamingHelper, i18nService, progressLedgerService);
-            FinalAnswerNode finalAnswerNode = new FinalAnswerNode(generatedFileCache);
+            FinalAnswerNode finalAnswerNode = new FinalAnswerNode(generatedFileCache, markdownNormalizeEnabled);
 
             KeyStrategyFactory keyStrategyFactory = KeyStrategy.builder()
                     // 输入字段

mateclaw-server/src/main/java/vip/mate/agent/graph/node/FinalAnswerNode.java

@@ -40,12 +40,25 @@ public class FinalAnswerNode implements NodeAction {
      */
     private final GeneratedFileCache generatedFileCache;
 
+    /**
+     * Kill-switch for the deterministic Markdown cleanup applied to the answer
+     * body. {@code true} (default) runs {@link MarkdownNormalizer}; set to
+     * {@code false} to surface model output verbatim if a normalization edge
+     * case ever mangles a legitimate answer in production.
+     */
+    private final boolean markdownNormalizeEnabled;
+
     public FinalAnswerNode() {
-        this(null);
+        this(null, true);
     }
 
     public FinalAnswerNode(GeneratedFileCache generatedFileCache) {
+        this(generatedFileCache, true);
+    }
+
+    public FinalAnswerNode(GeneratedFileCache generatedFileCache, boolean markdownNormalizeEnabled) {
         this.generatedFileCache = generatedFileCache;
+        this.markdownNormalizeEnabled = markdownNormalizeEnabled;
     }
 
     @Override
@@ -176,8 +189,11 @@ public class FinalAnswerNode implements NodeAction {
         // unaligned table pipes) that prompt rules fail to prevent; this fixes the
         // mechanical defects before the answer is persisted / sent to channels.
         // Verbatim tool output (RETURN_DIRECT) and approval-wait paths return early
-        // above and are intentionally left untouched.
+        // above and are intentionally left untouched. Gated so operators can turn
-        finalAnswer = MarkdownNormalizer.normalize(finalAnswer);
+        // the rewrite off (mate.agent.markdown-normalize-enabled=false).
+        if (markdownNormalizeEnabled) {
+            finalAnswer = MarkdownNormalizer.normalize(finalAnswer);
+        }
 
         // Build the event list. Always carries the finish_reason event so
         // downstream consumers (memory gate, channel accumulator, message

mateclaw-server/src/main/resources/application.yml

@@ -222,6 +222,10 @@ mateclaw:
 # MateClaw Agent 配置
 mate:
   agent:
+    # Deterministic Markdown cleanup of the final answer (heading spaces, glued
+    # ---, table pipe alignment) before persistence / channel delivery. Set to
+    # false to pass model output through verbatim.
+    markdown-normalize-enabled: true
     graph:
       observation:
         # 与 GraphObservationProperties.java 默认值对齐，参考 openclaw token-budget 设计