fix(channel): evict SSE RunState by inactivity, not wall-clock age

mateclaw-server/src/main/java/vip/mate/channel/web/ChatStreamTracker.java

@@ -1464,17 +1464,64 @@ public class ChatStreamTracker {
 
     /** 已完成的 RunState 保留时间（5 分钟） */
     private static final long DONE_RETENTION_MS = 5 * 60 * 1000;
-    /** RunState 最大存活时间（30 分钟，防止挂起的流永远占内存） */
+
-    private static final long MAX_LIFETIME_MS = 30 * 60 * 1000;
+    /**
+     * RunState 最长无活动时间。从 wall-clock {@code MAX_LIFETIME_MS=30min}
+     * 切换到 inactivity-based 后默认 30 min — 与 hermes-agent 的
+     * {@code gateway_timeout=1800s} 同口径：只要 agent 还在持续产事件
+     * （tool call / content delta / phase transition / progress_update），
+     * 就一直活下去，墙钟跑 1 小时 2 小时都可以。只有真正"完全静默 ≥ N 分钟"
+     * 才视为卡死并强制清理。
+     *
+     * <p>修复的背景：round-6 的 10-LLM 横评任务实际跑了 47 min，全程都在
+     * 出 tool call，但旧的 wall-clock 30 min 死线在 iter 128 / 8 of 10
+     * 就把 RunState 清掉了 — SSE 流死、UI 空白、用户以为任务挂了。换成
+     * inactivity 后，那种长任务永远不会被误清，而真正卡死的 agent（无活动
+     * 5+ 分钟）会按时清理。可通过 property
+     * {@code mateclaw.sse.idle-timeout-minutes} 调整。
+     */
+    @org.springframework.beans.factory.annotation.Value("${mateclaw.sse.idle-timeout-minutes:30}")
+    private int idleTimeoutMinutes = 30;
+
+    /**
+     * Test hook — backdates the {@code lastEventAt} timestamp on an
+     * existing RunState so {@link #cleanupStaleRuns()} can be exercised
+     * deterministically without sleeping for minutes. Package-private on
+     * purpose; production callers go through {@link #broadcast} which
+     * stamps the field forward.
+     */
+    void backdateLastEventForTesting(String conversationId, long lastEventAt) {
+        RunState state = runs.get(conversationId);
+        if (state != null) {
+            state.lastEventAt = lastEventAt;
+        }
+    }
+
+    /** Test hook — true when a RunState row exists for the conversation. */
+    boolean hasRunStateForTesting(String conversationId) {
+        return runs.containsKey(conversationId);
+    }
+
+    /** Test hook — exposes the configurable timeout for assertion. */
+    int idleTimeoutMinutesForTesting() {
+        return idleTimeoutMinutes;
+    }
+
+    /** Test hook — override the timeout in pure-unit tests that bypass Spring. */
+    void setIdleTimeoutMinutesForTesting(int minutes) {
+        this.idleTimeoutMinutes = minutes;
+    }
 
     /**
      * 定期清理过期的 RunState，防止内存泄漏。
-     * - 已完成超过 5 分钟的 → 移除
+     * - 已完成超过 {@link #DONE_RETENTION_MS} 的 → 移除
-     * - 存活超过 30 分钟的（无论是否完成）→ 强制移除
+     * - 自 {@link RunState#lastEventAt} 算起静默超过
+     *   {@link #idleTimeoutMinutes} 分钟的 → 强制移除（视为卡死）
      */
     @org.springframework.scheduling.annotation.Scheduled(fixedRate = 600_000)
     public void cleanupStaleRuns() {
         long now = System.currentTimeMillis();
+        long idleThresholdMs = (long) idleTimeoutMinutes * 60_000L;
         int evicted = 0;
 
         var iterator = runs.entrySet().iterator();
@@ -1482,6 +1529,7 @@ public class ChatStreamTracker {
             var entry = iterator.next();
             RunState state = entry.getValue();
             long age = now - state.createdAt;
+            long idleMs = now - state.lastEventAt;
 
             boolean shouldEvict = false;
             String reason = null;
@@ -1489,9 +1537,11 @@ public class ChatStreamTracker {
             if (state.done && age > DONE_RETENTION_MS) {
                 shouldEvict = true;
                 reason = "completed and expired";
-            } else if (age > MAX_LIFETIME_MS) {
+            } else if (idleMs > idleThresholdMs) {
                 shouldEvict = true;
-                reason = "exceeded max lifetime (" + (age / 1000) + "s)";
+                reason = "idle for " + (idleMs / 1000) + "s (threshold "
+                        + idleTimeoutMinutes + "min); total wall-clock age "
+                        + (age / 1000) + "s";
             }
 
             if (shouldEvict) {

mateclaw-server/src/test/java/vip/mate/channel/web/ChatStreamTrackerCleanupTest.java

@@ -0,0 +1,86 @@
+package vip.mate.channel.web;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Pins {@link ChatStreamTracker#cleanupStaleRuns()} — the switch from
+ * wall-clock {@code MAX_LIFETIME_MS} to inactivity-based eviction.
+ *
+ * <p>Before the fix, a long-running agent that kept producing tool calls
+ * (47-minute LLM-review smoke test, round 6) was killed at the 30-minute
+ * wall-clock mark mid-task. The new behaviour mirrors hermes-agent's
+ * {@code gateway_timeout}: only completely idle runs are evicted, the
+ * actively-producing ones can run as long as they need to.
+ */
+class ChatStreamTrackerCleanupTest {
+
+    private static ChatStreamTracker newTracker() {
+        ChatStreamTracker t = new ChatStreamTracker(new ObjectMapper());
+        // Tighten the idle threshold so the test stays at unit-test speed.
+        t.setIdleTimeoutMinutesForTesting(5);
+        return t;
+    }
+
+    @Test
+    @DisplayName("Active run (recent lastEventAt) survives cleanup regardless of total age.")
+    void activeRunSurvives() {
+        ChatStreamTracker tracker = newTracker();
+        tracker.register("conv-active");
+        // lastEventAt was set to "now" inside the RunState constructor —
+        // no backdate, so even a "very old createdAt" would be irrelevant.
+        tracker.cleanupStaleRuns();
+        assertTrue(tracker.hasRunStateForTesting("conv-active"),
+                "actively-producing run must not be evicted");
+    }
+
+    @Test
+    @DisplayName("Idle run beyond threshold is evicted.")
+    void idleRunEvicted() {
+        ChatStreamTracker tracker = newTracker();
+        tracker.register("conv-idle");
+        // 6 minutes ago — past the 5-minute threshold set above.
+        tracker.backdateLastEventForTesting("conv-idle", System.currentTimeMillis() - 6 * 60_000L);
+        tracker.cleanupStaleRuns();
+        assertFalse(tracker.hasRunStateForTesting("conv-idle"),
+                "idle run past threshold must be evicted");
+    }
+
+    @Test
+    @DisplayName("Idle run just inside the threshold survives — no premature eviction.")
+    void idleRunWithinThresholdSurvives() {
+        ChatStreamTracker tracker = newTracker();
+        tracker.register("conv-borderline");
+        // 4 minutes idle — within 5-minute window.
+        tracker.backdateLastEventForTesting("conv-borderline", System.currentTimeMillis() - 4 * 60_000L);
+        tracker.cleanupStaleRuns();
+        assertTrue(tracker.hasRunStateForTesting("conv-borderline"),
+                "run idle below threshold must survive — would otherwise be a regression of the wall-clock bug");
+    }
+
+    @Test
+    @DisplayName("Mixed: active + idle runs — only the idle one is evicted.")
+    void mixedRunsSelectiveEviction() {
+        ChatStreamTracker tracker = newTracker();
+        tracker.register("conv-active");
+        tracker.register("conv-idle");
+        tracker.backdateLastEventForTesting("conv-idle", System.currentTimeMillis() - 10 * 60_000L);
+        tracker.cleanupStaleRuns();
+        assertTrue(tracker.hasRunStateForTesting("conv-active"));
+        assertFalse(tracker.hasRunStateForTesting("conv-idle"));
+    }
+
+    @Test
+    @DisplayName("Default idle timeout from @Value matches the documented 30-minute fallback.")
+    void defaultIdleTimeoutIs30() {
+        // Bypass the @Value injection (no Spring context in this unit test) —
+        // the field initialiser pins the default so a refactor that drops the
+        // = 30 falls over here.
+        ChatStreamTracker tracker = new ChatStreamTracker(new ObjectMapper());
+        org.junit.jupiter.api.Assertions.assertEquals(30, tracker.idleTimeoutMinutesForTesting());
+    }
+}