feat(context): preserve spill markers across compaction phases, persist enriched boundary, broadcast compact_status SSE (#110)

mateclaw-server/src/main/java/vip/mate/agent/context/ConversationWindowManager.java

@@ -22,6 +22,7 @@ import vip.mate.workspace.conversation.ConversationService;
 
 import java.util.ArrayList;
 import java.util.List;
+import java.util.Map;
 import java.util.concurrent.ConcurrentHashMap;
 
 /**
@@ -131,6 +132,21 @@ public class ConversationWindowManager {
         this.toolResultStorage = toolResultStorage;
     }
 
+    /**
+     * Optional stream tracker for broadcasting {@code compact_status}
+     * SSE events. Wired via setter so unit tests can leave it {@code null}
+     * without dragging in the channel layer. When present, every
+     * compaction emits start/skipped/summarize/done events so the
+     * frontend can render a boundary card and a status line in real
+     * time.
+     */
+    private vip.mate.channel.web.ChatStreamTracker streamTracker;
+
+    @org.springframework.beans.factory.annotation.Autowired(required = false)
+    public void setStreamTracker(vip.mate.channel.web.ChatStreamTracker streamTracker) {
+        this.streamTracker = streamTracker;
+    }
+
     // ==================== 状态 ====================
 
     /** 摘要缓存：key = "conversationId:oldMessageCount" */
@@ -205,6 +221,8 @@ public class ConversationWindowManager {
         if (messages == null || messages.isEmpty()) {
             return messages;
         }
+        long spillsAtEntry = (toolResultStorage != null) ? toolResultStorage.getSpillCount() : 0L;
+
         messages = pruneOldToolResultsForModelInput(messages, conversationId, workspaceBasePath);
 
         int effectiveMax = (maxInputTokens != null && maxInputTokens > 0)
@@ -229,7 +247,7 @@ public class ConversationWindowManager {
 
         // 可用于历史的 token 预算 = max - system - currentMsg - tools - 安全余量
         int reservedTokens = systemTokens + currentMsgTokens + toolsTokens + (int) (effectiveMax * 0.05);
-        // RFC-025 Change 1: reserve 硬封顶到 effectiveMax 的 50%。
+        // 预留 reserve 硬封顶到 effectiveMax 的 50%。
         // 小上下文模型（Ollama 16K、本地 8K）下，systemTokens + currentMsgTokens 很容易
         // 接近或超过 effectiveMax，不封顶会让 historyBudget 变负数导致死循环压缩
         // （压缩目标比压缩前还大 → 压缩后又触发压缩）。
@@ -244,7 +262,8 @@ public class ConversationWindowManager {
         // 尾部保护 token 预算：阈值的 20%（与 Hermes 一致）
         int tailTokenBudget = (int) (triggerThreshold * 0.20);
 
-        return compactMessages(messages, historyBudget, tailTokenBudget, chatModel, conversationId, agentId);
+        return compactMessages(messages, historyBudget, tailTokenBudget, chatModel,
+                conversationId, agentId, totalTokens, spillsAtEntry);
     }
 
     /**
@@ -260,15 +279,40 @@ public class ConversationWindowManager {
 
     // ==================== 核心压缩逻辑 ====================
 
+    /** Broadcast a single compact_status event; silent no-op when no tracker is wired. */
+    private void broadcastCompactStatus(String conversationId, String status, Map<String, Object> extra) {
+        if (streamTracker == null || conversationId == null || conversationId.isEmpty()) {
+            return;
+        }
+        try {
+            Map<String, Object> payload = new java.util.LinkedHashMap<>();
+            payload.put("status", status);
+            payload.put("timestamp", System.currentTimeMillis());
+            if (extra != null) payload.putAll(extra);
+            streamTracker.broadcastObject(conversationId, "compact_status", payload);
+        } catch (Exception e) {
+            log.debug("[ConversationWindow] broadcast compact_status failed: {}", e.getMessage());
+        }
+    }
+
     private List<Message> compactMessages(List<Message> messages, int historyBudget,
                                           int tailTokenBudget, ChatModel chatModel,
-                                          String conversationId, Long agentId) {
+                                          String conversationId, Long agentId,
+                                          int preTokens, long spillsAtEntry) {
+        broadcastCompactStatus(conversationId, "start", Map.of(
+                "preTokens", preTokens,
+                "messagesIn", messages.size(),
+                "trigger", "token_threshold"
+        ));
+
         // 动态计算尾部保护边界（替代固定 preserveRecentPairs）
         int headEnd = 0; // 头部保护：暂不保护（system prompt 已在外部计算）
         int tailStart = findTailBoundary(messages, headEnd, tailTokenBudget);
 
         if (tailStart <= headEnd) {
             log.debug("[ConversationWindow] 消息数不足以拆分，跳过压缩");
+            broadcastCompactStatus(conversationId, "skipped",
+                    Map.of("reason", "insufficient_messages"));
             return messages;
         }
 
@@ -281,8 +325,14 @@ public class ConversationWindowManager {
         // turn.
         int pairSafeCut = enforcePairSafeBoundary(messages, headEnd, tailStart);
         if (pairSafeCut <= headEnd) {
+            broadcastCompactStatus(conversationId, "skipped",
+                    Map.of("reason", "pair_boundary_collapsed"));
             return messages;
         }
+        if (pairSafeCut != tailStart) {
+            broadcastCompactStatus(conversationId, "pair_safe", Map.of(
+                    "movedFrom", tailStart, "movedTo", pairSafeCut));
+        }
         tailStart = pairSafeCut;
 
         List<Message> oldMessages = new ArrayList<>(messages.subList(headEnd, tailStart));
@@ -340,13 +390,20 @@ public class ConversationWindowManager {
         // 计算动态摘要预算
         int summaryBudget = computeSummaryBudget(forSummary);
 
+        broadcastCompactStatus(conversationId, "summarize", Map.of(
+                "messagesToSummarize", oldMessages.size(),
+                "summaryBudget", summaryBudget
+        ));
+
         // 检查缓存
         String cacheKey = conversationId + ":" + oldMessages.size();
         CachedSummary cached = summaryCache.get(cacheKey);
         String summary;
+        boolean fromCache = false;
 
         if (cached != null && !cached.isExpired(CACHE_TTL_MS)) {
             summary = cached.summary();
+            fromCache = true;
             log.debug("[ConversationWindow] 命中摘要缓存, conv={}", conversationId);
         } else {
             summary = generateSummary(forSummary, chatModel, conversationId, summaryBudget, memoryExtraContext);
@@ -355,21 +412,12 @@ public class ConversationWindowManager {
                 int count = compressionCounts.merge(conversationId, 1, Integer::sum);
                 log.info("[ConversationWindow] 生成结构化摘要 ({} 字符, 第 {} 次压缩), 压缩 {} 条旧消息, conv={}",
                         summary.length(), count, oldMessages.size(), conversationId);
-
-                // 持久化摘要到 DB：下次加载历史时可直接从摘要位置开始，跳过重复压缩
-                if (conversationService != null) {
-                    try {
-                        conversationService.saveCompressionSummary(
-                                conversationId, SUMMARY_PREFIX + summary, oldMessages.size());
-                    } catch (Exception e) {
-                        log.warn("[ConversationWindow] Failed to persist compression summary: {}", e.getMessage());
-                    }
-                }
             }
         }
 
         // 组装结果
         List<Message> result = new ArrayList<>();
+        boolean anchored = false;
         if (summary != null && !summary.isBlank()) {
             result.add(new UserMessage(SUMMARY_PREFIX + summary));
 
@@ -380,11 +428,16 @@ public class ConversationWindowManager {
             Message anchor = buildFirstUserAnchor(oldMessages);
             if (anchor != null) {
                 result.add(anchor);
+                anchored = true;
             }
         } else if (!oldMessages.isEmpty()) {
             log.warn("[ConversationWindow] 摘要生成失败，降级为保留最近 4 条旧消息, conv={}", conversationId);
             int fallbackKeep = Math.min(4, oldMessages.size());
             result.addAll(oldMessages.subList(oldMessages.size() - fallbackKeep, oldMessages.size()));
+            broadcastCompactStatus(conversationId, "failed", Map.of(
+                    "reason", "summary_generation_failed",
+                    "fallbackKept", fallbackKeep
+            ));
         }
         result.addAll(recentMessages);
 
@@ -393,6 +446,42 @@ public class ConversationWindowManager {
         if (resultTokens > historyBudget && result.size() > 2) {
             log.warn("[ConversationWindow] 压缩后仍超预算: {} > {}, 执行二次裁剪", resultTokens, historyBudget);
             result = trimToFit(result, historyBudget);
+            resultTokens = TokenEstimator.estimateTokens(result);
+        }
+
+        // Persist the boundary + announce completion only when the summary
+        // actually wrote a row. Failed-summary fallback already broadcast
+        // its own event above.
+        if (summary != null && !summary.isBlank() && conversationService != null && !fromCache) {
+            long spillsThisTurn = (toolResultStorage != null)
+                    ? Math.max(0L, toolResultStorage.getSpillCount() - spillsAtEntry)
+                    : 0L;
+            Map<String, Object> boundaryMetadata = new java.util.LinkedHashMap<>();
+            boundaryMetadata.put("trigger", "token_threshold");
+            boundaryMetadata.put("preTokens", preTokens);
+            boundaryMetadata.put("postTokens", resultTokens);
+            boundaryMetadata.put("messagesSummarized", oldMessages.size());
+            boundaryMetadata.put("tailKept", recentMessages.size());
+            boundaryMetadata.put("toolResultsSpilled", spillsThisTurn);
+            boundaryMetadata.put("anchored", anchored);
+            try {
+                conversationService.saveCompressionSummary(
+                        conversationId, SUMMARY_PREFIX + summary, oldMessages.size(),
+                        boundaryMetadata);
+            } catch (Exception e) {
+                log.warn("[ConversationWindow] Failed to persist compression boundary: {}", e.getMessage());
+            }
+            broadcastCompactStatus(conversationId, "done", boundaryMetadata);
+        } else if (summary != null && !summary.isBlank() && fromCache) {
+            // Cached summary path — no new DB row, but emit done so the
+            // frontend status bar still updates.
+            broadcastCompactStatus(conversationId, "done", Map.of(
+                    "preTokens", preTokens,
+                    "postTokens", resultTokens,
+                    "messagesSummarized", oldMessages.size(),
+                    "tailKept", recentMessages.size(),
+                    "fromCache", true
+            ));
         }
 
         return result;
@@ -780,15 +869,36 @@ public class ConversationWindowManager {
     }
 
     /**
-     * Phase 1 - Soft trim：对工具结果做 head+tail 裁剪（保留首尾各 200 字符）。
+     * Spill-marker responses already point at an on-disk full copy via
+     * {@code path=...} in their body. Trimming, replacing, or pre-pruning
+     * them would destroy the very pointer the model needs to recover the
+     * original output with {@code read_file} — which is the whole reason
+     * we spilled in the first place. All three compaction phases consult
+     * this guard before touching a response.
      */
-    private int softTrimToolResults(List<Message> messages) {
+    static boolean isSpillMarker(ToolResponseMessage.ToolResponse r) {
+        return r != null
+                && r.responseData() != null
+                && r.responseData().startsWith(ToolResultStorage.SPILL_MARKER_PREFIX);
+    }
+
+    /**
+     * Phase 1 - Soft trim：对工具结果做 head+tail 裁剪（保留首尾各 200 字符）。
+     * <p>Spill-marker responses are left untouched so their on-disk pointer
+     * survives intact across compaction.
+     */
+    int softTrimToolResults(List<Message> messages) {
         int trimmed = 0;
         for (int i = 0; i < messages.size(); i++) {
             if (messages.get(i) instanceof ToolResponseMessage trm) {
                 List<ToolResponseMessage.ToolResponse> newResponses = new ArrayList<>();
                 boolean changed = false;
                 for (ToolResponseMessage.ToolResponse r : trm.getResponses()) {
+                    if (isSpillMarker(r)) {
+                        // Pointer + preview already; trimming would lose the path.
+                        newResponses.add(r);
+                        continue;
+                    }
                     String data = r.responseData();
                     if (data != null && data.length() > 500) {
                         String head = data.substring(0, 200);
@@ -811,16 +921,28 @@ public class ConversationWindowManager {
 
     /**
      * Phase 2 - Hard clear：将所有旧工具结果替换为占位符。
+     * <p>Spill-marker responses are left untouched so the on-disk pointer
+     * survives — a placeholder here would force the model to abandon a
+     * tool output it could otherwise recover via {@code read_file}.
      */
-    private int hardClearToolResults(List<Message> messages) {
+    int hardClearToolResults(List<Message> messages) {
         int cleared = 0;
         for (int i = 0; i < messages.size(); i++) {
             if (messages.get(i) instanceof ToolResponseMessage trm) {
-                List<ToolResponseMessage.ToolResponse> placeholders = trm.getResponses().stream()
-                        .map(r -> new ToolResponseMessage.ToolResponse(r.id(), r.name(), "[tool result removed]"))
-                        .toList();
-                messages.set(i, ToolResponseMessage.builder().responses(placeholders).build());
-                cleared++;
+                boolean changed = false;
+                List<ToolResponseMessage.ToolResponse> replaced = new ArrayList<>(trm.getResponses().size());
+                for (ToolResponseMessage.ToolResponse r : trm.getResponses()) {
+                    if (isSpillMarker(r)) {
+                        replaced.add(r);
+                        continue;
+                    }
+                    replaced.add(new ToolResponseMessage.ToolResponse(r.id(), r.name(), "[tool result removed]"));
+                    changed = true;
+                }
+                if (changed) {
+                    messages.set(i, ToolResponseMessage.builder().responses(replaced).build());
+                    cleared++;
+                }
             }
         }
         return cleared;
@@ -828,18 +950,27 @@ public class ConversationWindowManager {
 
     /**
      * Phase 3 Pre-prune：在 LLM 摘要前，将工具输出替换为占位符（减少摘要输入 token）。
+     * <p>Spill-marker responses are left untouched so the summary input
+     * still has the on-disk path the model might cite back in its summary.
      */
-    private int prePruneForSummary(List<Message> messages) {
+    int prePruneForSummary(List<Message> messages) {
         int pruned = 0;
         for (int i = 0; i < messages.size(); i++) {
             if (messages.get(i) instanceof ToolResponseMessage trm) {
                 boolean hasSubstantial = trm.getResponses().stream()
-                        .anyMatch(r -> r.responseData() != null && r.responseData().length() > 200);
+                        .anyMatch(r -> !isSpillMarker(r)
+                                && r.responseData() != null
+                                && r.responseData().length() > 200);
                 if (hasSubstantial) {
-                    List<ToolResponseMessage.ToolResponse> placeholders = trm.getResponses().stream()
-                            .map(r -> new ToolResponseMessage.ToolResponse(r.id(), r.name(),
-                                    "[旧工具输出已清理以节省上下文空间]"))
-                            .toList();
+                    List<ToolResponseMessage.ToolResponse> placeholders = new ArrayList<>(trm.getResponses().size());
+                    for (ToolResponseMessage.ToolResponse r : trm.getResponses()) {
+                        if (isSpillMarker(r)) {
+                            placeholders.add(r);
+                            continue;
+                        }
+                        placeholders.add(new ToolResponseMessage.ToolResponse(r.id(), r.name(),
+                                "[旧工具输出已清理以节省上下文空间]"));
+                    }
                     messages.set(i, ToolResponseMessage.builder().responses(placeholders).build());
                     pruned++;
                 }

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

@@ -99,12 +99,17 @@ public class ToolResultProperties {
 
     /**
      * Days to retain spill files before the scheduled cleanup deletes them.
-     * Spill files exist to let the model recover full tool output via
-     * {@code read_file} during the active conversation; once the conversation
-     * is dormant for this many days the agent is extremely unlikely to ever
-     * read the file again, and disk pressure starts to matter.
+     * <p><b>Default 0 means time-based cleanup is disabled</b> — spill files
+     * stay on disk until the owning conversation is explicitly deleted (which
+     * fires {@code purgeConversation} via {@code ConversationService}).
+     * This preserves the "recoverable" invariant: a summary or preview that
+     * cites a spill path will keep working for the whole life of the
+     * conversation, no matter how long it sits dormant.
+     * <p>Set to a positive value if disk pressure outweighs recoverability
+     * for your deployment. The scheduled sweep will then delete files whose
+     * mtime falls outside the retention horizon.
      */
-    private int retentionDays = 7;
+    private int retentionDays = 0;
 
     /**
      * Cron expression for the spill-cleanup task. Defaults to once a day at

mateclaw-server/src/main/java/vip/mate/workspace/conversation/ConversationService.java

@@ -428,18 +428,65 @@ public class ConversationService {
     }
 
     /**
-     * 将压缩摘要持久化为 role=system 的特殊消息。
-     * 下次加载历史时识别此消息，跳过它之前的已压缩消息。
+     * Persist a compaction boundary as a role=system message. The body is
+     * the summary text; the metadata describes <em>what happened</em> at
+     * this boundary (trigger, pre/post tokens, how many messages were
+     * summarised, how many spill files were produced, how many tail
+     * messages survived). On the next load this row is the cut-off — older
+     * messages are skipped, the model picks up from the summary forward.
+     *
+     * <p>Backward-compat overload: legacy callers that only know the row
+     * count still work and produce a minimal metadata block.
      */
     public void saveCompressionSummary(String conversationId, String summary, int compressedCount) {
+        saveCompressionSummary(conversationId, summary, compressedCount, Map.of());
+    }
+
+    /**
+     * Same as the 3-arg overload but accepts extra structured fields that
+     * are merged into the boundary's metadata JSON. Fields the frontend
+     * and observability pipeline care about:
+     * <ul>
+     *   <li>{@code trigger} — what fired this boundary
+     *       ({@code token_threshold}, {@code user_compact}, etc.)</li>
+     *   <li>{@code preTokens} / {@code postTokens} — context size before
+     *       and after, for the in-prompt status row</li>
+     *   <li>{@code messagesSummarized} / {@code tailKept} — partition
+     *       counts the user sees in the boundary card</li>
+     *   <li>{@code toolResultsSpilled} — how many bodies the spill store
+     *       absorbed during this boundary</li>
+     *   <li>{@code summaryId} — stable id (the inserted message id) for
+     *       deep-linking from the SSE event</li>
+     * </ul>
+     * <p>{@code type=compression_summary} is always present — the loader
+     * keys off it. {@code compressedCount} is kept for backward compat.
+     */
+    public void saveCompressionSummary(String conversationId, String summary, int compressedCount,
+                                       Map<String, Object> extraMetadata) {
         MessageEntity entity = new MessageEntity();
         entity.setConversationId(conversationId);
         entity.setRole("system");
         entity.setContent(summary);
         entity.setStatus("completed");
-        entity.setMetadata("{\"type\":\"compression_summary\",\"compressedCount\":" + compressedCount + "}");
+
+        Map<String, Object> metadata = new java.util.LinkedHashMap<>();
+        metadata.put("type", "compression_summary");
+        metadata.put("compressedCount", compressedCount);
+        if (extraMetadata != null) {
+            extraMetadata.forEach((k, v) -> {
+                if (v != null) metadata.put(k, v);
+            });
+        }
+        try {
+            entity.setMetadata(objectMapper.writeValueAsString(metadata));
+        } catch (com.fasterxml.jackson.core.JsonProcessingException e) {
+            log.warn("[Conversation] Failed to serialise compaction metadata, falling back to minimal: {}",
+                    e.getMessage());
+            entity.setMetadata("{\"type\":\"compression_summary\",\"compressedCount\":" + compressedCount + "}");
+        }
         messageMapper.insert(entity);
-        log.info("[Conversation] Saved compression summary for conv={}, compressedCount={}", conversationId, compressedCount);
+        log.info("[Conversation] Saved compression boundary conv={}, compressedCount={}, metadata={}",
+                conversationId, compressedCount, entity.getMetadata());
     }
 
     public List<MessageVO> listMessageViews(String conversationId) {

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

@@ -229,10 +229,13 @@ mate:
       excluded-tools:
         - read_file
         - read_workspace_memory_file
-      # Spill files are deleted after this many days. Set to 0 to disable the
-      # scheduled sweep entirely (files still get purged when the conversation
-      # is deleted explicitly via ConversationService.deleteConversation).
-      retention-days: 7
+      # Spill files are deleted after this many days. Default 0 disables the
+      # scheduled sweep entirely so a summary/preview that points at a spill
+      # path stays valid for the whole life of the conversation. Files are
+      # still purged when the conversation is deleted explicitly via
+      # ConversationService.deleteConversation. Raise to a positive value if
+      # disk pressure outweighs recoverability for your deployment.
+      retention-days: 0
       cleanup-cron: "0 0 3 * * ?"
     conversation:
       window:

mateclaw-server/src/test/java/vip/mate/agent/context/ConversationWindowManagerSpillMarkerPreservationTest.java

@@ -0,0 +1,157 @@
+package vip.mate.agent.context;
+
+import org.junit.jupiter.api.Test;
+import org.springframework.ai.chat.messages.Message;
+import org.springframework.ai.chat.messages.ToolResponseMessage;
+import vip.mate.agent.graph.executor.ToolResultStorage;
+import vip.mate.config.ConversationWindowProperties;
+
+import java.util.ArrayList;
+import java.util.List;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * The three compaction phases (soft trim, hard clear, pre-prune for
+ * summary) must never destroy a spill-marker body — doing so would erase
+ * the {@code path=...} pointer the model needs to recover the original
+ * full output via {@code read_file}, which is the whole reason that body
+ * was spilled in the first place.
+ *
+ * <p>This is the "recoverable" invariant: once a tool output makes it
+ * into the spill store, the in-context representation stays a stable
+ * preview + path for the rest of the conversation regardless of how
+ * aggressively the window manager has to compress the prefix.
+ */
+class ConversationWindowManagerSpillMarkerPreservationTest {
+
+    private static final String SPILL_BODY = ToolResultStorage.SPILL_MARKER_PREFIX
+            + " tool=web_search full_chars=22000 path=/tmp/x.txt\n"
+            + "[Preview — first 800 of 22000 chars. Use read_file with the path above to retrieve the rest.]\n"
+            + "preview body fragment that contributes most of the inline size...";
+
+    @Test
+    void softTrimLeavesSpillMarkerUntouched() {
+        ConversationWindowManager mgr = new ConversationWindowManager(
+                new ConversationWindowProperties(), null, null);
+
+        List<Message> messages = new ArrayList<>(List.of(
+                toolMessage("call-spill", "web_search", SPILL_BODY),
+                toolMessage("call-big", "search", "x".repeat(2000))
+        ));
+
+        // Pass the same list through Phase 1.
+        int trimmed = mgr.softTrimToolResults(messages);
+
+        // The non-spill body must have been trimmed (it was > 500 chars).
+        // The spill body must remain identical to the original.
+        ToolResponseMessage trm0 = (ToolResponseMessage) messages.get(0);
+        ToolResponseMessage trm1 = (ToolResponseMessage) messages.get(1);
+        assertEquals(SPILL_BODY, trm0.getResponses().getFirst().responseData(),
+                "Phase 1 soft trim must not modify a spill-marker body");
+        assertTrue(trm1.getResponses().getFirst().responseData().contains("[trimmed "),
+                "non-spill bodies should still be trimmed by Phase 1");
+        assertEquals(1, trimmed,
+                "trim counter should reflect only the non-spill body that was actually shortened");
+    }
+
+    @Test
+    void hardClearLeavesSpillMarkerUntouched() {
+        ConversationWindowManager mgr = new ConversationWindowManager(
+                new ConversationWindowProperties(), null, null);
+
+        List<Message> messages = new ArrayList<>(List.of(
+                toolMessage("call-spill", "web_search", SPILL_BODY),
+                toolMessage("call-big", "search", "y".repeat(2000))
+        ));
+
+        int cleared = mgr.hardClearToolResults(messages);
+
+        ToolResponseMessage trm0 = (ToolResponseMessage) messages.get(0);
+        ToolResponseMessage trm1 = (ToolResponseMessage) messages.get(1);
+        assertEquals(SPILL_BODY, trm0.getResponses().getFirst().responseData(),
+                "Phase 2 hard clear must not replace a spill-marker body with [tool result removed]");
+        assertEquals("[tool result removed]", trm1.getResponses().getFirst().responseData(),
+                "non-spill bodies should still be replaced by Phase 2");
+        assertEquals(1, cleared,
+                "clear counter should reflect only the non-spill body that was actually replaced");
+    }
+
+    @Test
+    void prePruneForSummaryLeavesSpillMarkerUntouched() {
+        ConversationWindowManager mgr = new ConversationWindowManager(
+                new ConversationWindowProperties(), null, null);
+
+        List<Message> messages = new ArrayList<>(List.of(
+                toolMessage("call-spill", "web_search", SPILL_BODY),
+                toolMessage("call-big", "search", "z".repeat(2000))
+        ));
+
+        int pruned = mgr.prePruneForSummary(messages);
+
+        ToolResponseMessage trm0 = (ToolResponseMessage) messages.get(0);
+        ToolResponseMessage trm1 = (ToolResponseMessage) messages.get(1);
+        assertEquals(SPILL_BODY, trm0.getResponses().getFirst().responseData(),
+                "Phase 3 pre-prune must not replace a spill-marker body with the cleared-output placeholder");
+        assertTrue(trm1.getResponses().getFirst().responseData().contains("旧工具输出已清理"),
+                "non-spill bodies should still be replaced by Phase 3");
+        assertEquals(1, pruned);
+    }
+
+    @Test
+    void mixedMessageWithSpillAndNonSpillResponsesPreservesOnlyTheMarker() {
+        // A single ToolResponseMessage can hold multiple ToolResponses (one
+        // assistant tool_calls turn could ask for several tools at once).
+        // The phase guards must operate at the response level, not the
+        // message level — the spill response stays, the non-spill response
+        // gets the placeholder.
+        ConversationWindowManager mgr = new ConversationWindowManager(
+                new ConversationWindowProperties(), null, null);
+
+        ToolResponseMessage mixed = ToolResponseMessage.builder().responses(List.of(
+                new ToolResponseMessage.ToolResponse("call-spill", "web_search", SPILL_BODY),
+                new ToolResponseMessage.ToolResponse("call-big", "search", "q".repeat(2000))
+        )).build();
+        List<Message> messages = new ArrayList<>(List.of(mixed));
+
+        mgr.hardClearToolResults(messages);
+
+        ToolResponseMessage trm = (ToolResponseMessage) messages.getFirst();
+        assertEquals(SPILL_BODY, trm.getResponses().get(0).responseData(),
+                "the spill response in a mixed message must survive Phase 2");
+        assertEquals("[tool result removed]", trm.getResponses().get(1).responseData(),
+                "the non-spill response in a mixed message must still be cleared");
+    }
+
+    @Test
+    void smallSpillMarkerStillStaysVerbatim() {
+        // Edge case: even when the preview is short (under the 500-char
+        // soft-trim threshold), the marker check should still apply. This
+        // protects against future changes to the trim threshold.
+        ConversationWindowManager mgr = new ConversationWindowManager(
+                new ConversationWindowProperties(), null, null);
+
+        String tinySpill = ToolResultStorage.SPILL_MARKER_PREFIX
+                + " tool=test full_chars=600 path=/tmp/t.txt\n[tiny]";
+        List<Message> messages = new ArrayList<>(List.of(
+                toolMessage("call-1", "test", tinySpill)
+        ));
+
+        mgr.softTrimToolResults(messages);
+        mgr.hardClearToolResults(messages);
+        mgr.prePruneForSummary(messages);
+
+        ToolResponseMessage trm = (ToolResponseMessage) messages.getFirst();
+        assertEquals(tinySpill, trm.getResponses().getFirst().responseData(),
+                "the marker check is what protects the body — not the size of the preview");
+    }
+
+    // ------------------------------------------------------------------ helpers
+
+    private static ToolResponseMessage toolMessage(String id, String name, String data) {
+        return ToolResponseMessage.builder()
+                .responses(List.of(new ToolResponseMessage.ToolResponse(id, name, data)))
+                .build();
+    }
+}

mateclaw-server/src/test/java/vip/mate/agent/graph/executor/LaneDExecutorAndConfigTest.java

@@ -103,6 +103,20 @@ class LaneDExecutorAndConfigTest {
                     "storageBaseDir should still default to empty string");
         }
 
+        @Test
+        @DisplayName("retentionDays defaults to 0 so spill files outlive their conversation")
+        void retentionDaysDefaultsToZero() {
+            // The recoverability invariant: a summary or preview that cites
+            // a spill path must keep working for the whole life of the
+            // conversation. Time-based deletion is opt-in; operators with
+            // disk pressure can raise this value explicitly.
+            ToolResultProperties props = new ToolResultProperties();
+            assertEquals(0, props.getRetentionDays(),
+                    "retentionDays must default to 0 — time-based purge is opt-in to preserve recoverability");
+            assertTrue(props.getCleanupCron() != null && !props.getCleanupCron().isBlank(),
+                    "cleanupCron stays defined; it is a no-op while retentionDays=0");
+        }
+
         @Test
         @DisplayName("Per-result threshold matches the executor inline hard cap so spill and truncate share one ladder")
         void thresholdMatchesExecutorHardCap() throws Exception {