perf(agent): implement Lane D performance fixes

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

@@ -256,9 +256,23 @@ public class NodeStreamingChatHelper {
         }
     }
 
+    /**
+     * Broadcast a lightweight progress event so the frontend shows activity
+     * during silent LLM calls (e.g. triage). Sent as a "progress" SSE event.
+     */
+    public void broadcastProgress(String conversationId, String message) {
+        if (streamTracker == null || conversationId == null || conversationId.isEmpty()) {
+            return;
+        }
+        streamTracker.broadcastObject(conversationId, "progress",
+                Map.of("message", message != null ? message : ""));
+    }
+
     // ==================== 重试配置 ====================
 
     private static final int MAX_RETRIES = 5;
+    // For rate-limit and server-error: fail fast to failover chain.
+    private static final int MAX_RETRIES_RATE_LIMIT = 2;
     private static final long BACKOFF_BASE_MS = 3000;
     private static final long BACKOFF_CAP_MS = 60_000;
 
@@ -540,11 +554,22 @@ public class NodeStreamingChatHelper {
                 broadcastDelta(conversationId, "warning",
                         buildDeltaJson("⏱️ 请求频率受限，等待 " + (delay / 1000) + " 秒后重试（第 " + attempt + "/" + MAX_RETRIES + " 次）..."));
             }
-            try {
+            // Poll stop flag every 100ms so user Stop is honored mid-backoff.
-                Thread.sleep(delay);
+            long remaining = delay;
-            } catch (InterruptedException ie) {
+            while (remaining > 0) {
-                Thread.currentThread().interrupt();
+                if (streamTracker != null && streamTracker.isStopRequested(conversationId)) {
-                return buildErrorResult("LLM 调用被中断", conversationId, phase);
+                    log.info("[{}] Stop requested during backoff — aborting retry: conversationId={}",
+                            phase, conversationId);
+                    throw new CancellationException("Stream stopped by user");
+                }
+                long slice = Math.min(100, remaining);
+                try {
+                    Thread.sleep(slice);
+                } catch (InterruptedException ie) {
+                    Thread.currentThread().interrupt();
+                    return buildErrorResult("LLM 调用被中断", conversationId, phase);
+                }
+                remaining -= slice;
             }
         }
 
@@ -731,10 +756,12 @@ public class NodeStreamingChatHelper {
                         conversationId, phase, errorType);
             }
 
-            // Rate limit / Server error: 重试
+            // Rate limit / Server error: retry with reduced limit for rate-limit/server-error
-            if (attempt < MAX_RETRIES && (errorType == ErrorType.RATE_LIMIT || errorType == ErrorType.SERVER_ERROR)) {
+            boolean isRateLimitOrServerError = errorType == ErrorType.RATE_LIMIT || errorType == ErrorType.SERVER_ERROR;
+            int effectiveMaxRetries = isRateLimitOrServerError ? MAX_RETRIES_RATE_LIMIT : MAX_RETRIES;
+            if (attempt < effectiveMaxRetries && isRateLimitOrServerError) {
                 log.warn("[{}] Retryable error (attempt {}/{}, type={}): {}",
-                        phase, attempt, MAX_RETRIES, errorType, error.getMessage());
+                        phase, attempt, effectiveMaxRetries, errorType, error.getMessage());
                 return null;  // 返回 null 触发重试
             }
 

mateclaw-server/src/main/java/vip/mate/agent/graph/executor/ToolExecutionExecutor.java

@@ -38,13 +38,10 @@ import java.util.concurrent.*;
 public class ToolExecutionExecutor {
 
     private static final ObjectMapper OBJECT_MAPPER = new ObjectMapper();
-    private static final ExecutorService TOOL_EXECUTOR = Executors.newFixedThreadPool(
+    // JDK 21 virtual threads: no blocking stall for I/O-bound tools.
-            Math.max(4, Runtime.getRuntime().availableProcessors()),
+    // Each tool invocation gets its own lightweight carrier thread.
-            r -> {
+    private static final ExecutorService TOOL_EXECUTOR =
-                Thread t = new Thread(r, "tool-executor");
+            Executors.newVirtualThreadPerTaskExecutor();
-                t.setDaemon(true);
-                return t;
-            });
 
     /**
      * Legacy hardcoded unsafe set, kept as a fallback when no

mateclaw-server/src/main/java/vip/mate/agent/graph/executor/ToolResultProperties.java

@@ -26,8 +26,8 @@ import java.util.Set;
  *   agent:
  *     tool-result:
  *       enabled: true
- *       per-result-threshold-chars: 4000
+ *       per-result-threshold-chars: 16000
- *       per-turn-budget-chars: 16000
+ *       per-turn-budget-chars: 32000
  *       preview-head-chars: 800
  *       storage-base-dir:
  * </pre>
@@ -39,10 +39,10 @@ public class ToolResultProperties {
     private boolean enabled = true;
 
     /** A single tool result larger than this is spilled to disk. */
-    private int perResultThresholdChars = 4000;
+    private int perResultThresholdChars = 16000;  // was 4000 — prevents WebSearch spill-to-disk
 
     /** Aggregate cap on combined response size in one tool turn. */
-    private int perTurnBudgetChars = 16000;
+    private int perTurnBudgetChars = 32000;  // was 16000 — headroom for multi-tool turns
 
     /** Number of leading characters kept inline as a preview after spilling. */
     private int previewHeadChars = 800;

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

@@ -165,6 +165,12 @@ public class PlanGenerationNode implements NodeAction {
 
             Prompt prompt = new Prompt(promptMessages);
 
+            // Broadcast a lightweight progress token so the frontend shows activity
+            // during the silent triage call (typically 1-3 s).
+            if (streamingHelper != null) {
+                streamingHelper.broadcastProgress(conversationId, "分析中...");
+            }
+
             // Silent streaming call — structured JSON is parsed below; tokens are not forwarded to the client.
             NodeStreamingChatHelper.StreamResult result = streamingHelper.streamCallSilent(
                     chatModel, prompt, conversationId, "plan_generation");

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

@@ -167,8 +167,8 @@ mate:
     # the in-context preview points the agent at the spill file (read_file tool).
     tool-result:
       enabled: true
-      per-result-threshold-chars: 4000
+      per-result-threshold-chars: 16000   # was 4000 — prevents WebSearch spill-to-disk
-      per-turn-budget-chars: 16000
+      per-turn-budget-chars: 32000        # was 16000 — headroom for multi-tool turns
       preview-head-chars: 800
       storage-base-dir: ""
       # Retrieval-style tools that must NEVER be spilled. Spilling read_file's