fix(agent): stop stale ledger entries from suppressing repeated status queries

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

@@ -13,6 +13,7 @@ import vip.mate.agent.context.ChatOrigin;
 import vip.mate.agent.context.ChatOriginHolder;
 import vip.mate.agent.event.AgentLifecycleEvent;
 import vip.mate.agent.model.AgentEntity;
+import vip.mate.agent.progress.ProgressLedgerService;
 import vip.mate.agent.repository.AgentMapper;
 import vip.mate.exception.MateClawException;
 import vip.mate.llm.chatmodel.ThinkingLevelHolder;
@@ -69,6 +70,14 @@ public class AgentService {
     @Autowired(required = false)
     private vip.mate.agent.runtime.RunningConversationRegistry runningConversationRegistry;
 
+    /**
+     * Optional — clears leftover auto-recorded ledger entries when a new
+     * user turn starts. Field-injected so existing test constructors of
+     * {@code AgentService} don't need to supply it.
+     */
+    @Autowired(required = false)
+    private ProgressLedgerService progressLedgerService;
+
     /**
      * Runtime Agent instance cache. Keyed first by agentId, then by a model
      * key, so a conversation that pins a non-default model gets its own graph
@@ -254,6 +263,32 @@ public class AgentService {
 
     // ==================== 运行时入口 ====================
 
+    /**
+     * New-user-turn housekeeping: drop auto-recorded ledger entries left
+     * over from the previous turn. They mark past tool calls as DONE, and
+     * the ledger snapshot's "已完成的步骤不要重复执行" instruction would
+     * otherwise stop the agent from re-running status-query tools when the
+     * user repeats a question that needs fresh data.
+     *
+     * <p>Only the fresh-turn entries ({@code chat} / {@code chatStream} /
+     * {@code chatStructuredStream} / {@code execute}) call this. The
+     * approval-replay entries ({@code chatWithReplay*}) resume the SAME
+     * logical turn after a tool approval and must keep the safety net for
+     * work already done before the pause.
+     */
+    private void clearAutoRecordedForNewTurn(String conversationId) {
+        if (progressLedgerService == null || conversationId == null || conversationId.isBlank()) {
+            return;
+        }
+        try {
+            progressLedgerService.clearAutoRecorded(conversationId);
+        } catch (Exception e) {
+            // Ledger housekeeping must never block the chat itself.
+            log.warn("Failed to clear auto-recorded ledger entries for {}: {}",
+                    conversationId, e.getMessage());
+        }
+    }
+
     public String chat(Long agentId, String message, String conversationId) {
         return chat(agentId, message, conversationId, ChatOrigin.EMPTY);
     }
@@ -264,6 +299,7 @@ public class AgentService {
      * down to {@code @Tool} methods via Spring AI {@link org.springframework.ai.chat.model.ToolContext}.
      */
     public String chat(Long agentId, String message, String conversationId, ChatOrigin origin) {
+        clearAutoRecordedForNewTurn(conversationId);
         memoryRecallTracker.trackRecalls(agentId, message);
         BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOriginHolder.set(origin != null ? origin : ChatOrigin.EMPTY);
@@ -300,6 +336,7 @@ public class AgentService {
     }
 
     public Flux<String> chatStream(Long agentId, String message, String conversationId, ChatOrigin origin) {
+        clearAutoRecordedForNewTurn(conversationId);
         memoryRecallTracker.trackRecalls(agentId, message);
         BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         // Capture the origin into a request-scoped holder; cleared on Flux
@@ -336,6 +373,7 @@ public class AgentService {
     public Flux<StreamDelta> chatStructuredStream(Long agentId, String message, String conversationId,
                                                    String requesterId, String thinkingLevel,
                                                    ChatOrigin origin) {
+        clearAutoRecordedForNewTurn(conversationId);
         memoryRecallTracker.trackRecalls(agentId, message);
         BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
 
@@ -382,6 +420,7 @@ public class AgentService {
     }
 
     public String execute(Long agentId, String goal, String conversationId, ChatOrigin origin) {
+        clearAutoRecordedForNewTurn(conversationId);
         memoryRecallTracker.trackRecalls(agentId, goal);
         BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOriginHolder.set(origin != null ? origin : ChatOrigin.EMPTY);

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

@@ -56,12 +56,28 @@ public class ActionNode implements NodeAction {
 
     /**
      * Tools whose results should NOT be auto-recorded into the ledger.
-     * Meta-tools (load_skill, enable_tool, progress_update) either have
-     * their own ledger side-effects or are the ledger itself.
+     * Two groups:
+     * <ul>
+     *   <li><b>Meta-tools</b> (load_skill, enable_tool, progress_update,
+     *       skill helpers) — they either have their own ledger side-effects
+     *       or are the ledger itself.</li>
+     *   <li><b>Read-only / status-query tools</b> — querying live state is
+     *       not a task step that must not be repeated. Recording it as DONE
+     *       (with a frozen result excerpt in the note) pushes the model to
+     *       answer follow-up questions from stale output instead of
+     *       re-checking, because the snapshot instructs "已完成的步骤不要
+     *       重复执行".</li>
+     * </ul>
      */
     private static final Set<String> AUTO_RECORD_SKIP = Set.of(
             LOAD_SKILL_TOOL, ENABLE_TOOL, PROGRESS_UPDATE_TOOL,
-            "listAvailableSkills", "readSkillFile", "runSkillScript"
+            "listAvailableSkills", "readSkillFile", "runSkillScript",
+            // read-only / status-query tools
+            "read_file", "web_search",
+            "extract_document_text", "extract_pdf_text", "extract_docx_text",
+            "detect_file_type",
+            "getCurrentDateTime", "getCurrentDate", "getCurrentTime",
+            "listSubagents"
     );
 
     private final ToolExecutionExecutor executor;
@@ -256,8 +272,8 @@ public class ActionNode implements NodeAction {
      * avoid collisions between servers that expose tools with the same slug.
      * The display label uses the simplified slug for readability.
      */
-    private void autoRecordToolCalls(String conversationId,
-                                     List<ToolResponseMessage.ToolResponse> responses) {
+    void autoRecordToolCalls(String conversationId,
+                             List<ToolResponseMessage.ToolResponse> responses) {
         if (progressLedgerService == null || conversationId == null
                 || conversationId.isBlank() || responses == null || responses.isEmpty()) {
             return;

mateclaw-server/src/main/java/vip/mate/agent/progress/ProgressLedgerService.java

@@ -202,6 +202,39 @@ public class ProgressLedgerService {
         }
     }
 
+    /**
+     * Remove every auto-recorded entry ({@code auto_} key prefix) from the
+     * conversation's ledger. Called at the start of each new user turn.
+     *
+     * <p>Auto-recorded entries are a safety net against context trimming
+     * <b>within</b> one turn's tool loop. Letting them survive into the next
+     * user turn is harmful: the snapshot renders them as DONE alongside the
+     * "已完成的步骤不要重复执行" instruction, which stops the agent from
+     * re-running read-only / status-query tools when the user repeats a
+     * question that needs fresh data (e.g. "看下会议室有没有人"), and the
+     * frozen 120-char result note tempts it to answer from stale output.
+     *
+     * <p>LLM-authored regular entries and pinned skill constraints are
+     * untouched — multi-turn task tracking keeps working.
+     */
+    public void clearAutoRecorded(String conversationId) {
+        if (conversationId == null || conversationId.isBlank()) {
+            return;
+        }
+        ReentrantLock lock = upsertLocks.computeIfAbsent(conversationId, k -> new ReentrantLock());
+        lock.lock();
+        try {
+            LedgerWrapper wrapper = loadWrapper(conversationId);
+            boolean removed = wrapper.entries.keySet().removeIf(
+                    k -> k != null && k.startsWith(ProgressLedger.AUTO_RECORDED_PREFIX));
+            if (removed) {
+                persistWrapper(conversationId, wrapper);
+            }
+        } finally {
+            lock.unlock();
+        }
+    }
+
     /**
      * Auto-record a completed tool call as a ledger entry (B5). Uses the
      * {@link ProgressLedger#AUTO_RECORDED_PREFIX} on the key so the renderer

mateclaw-server/src/test/java/vip/mate/agent/graph/node/ActionNodeAutoRecordSkipTest.java

@@ -0,0 +1,117 @@
+package vip.mate.agent.graph.node;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.datatype.jsr310.JavaTimeModule;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+import org.springframework.ai.chat.messages.ToolResponseMessage;
+import vip.mate.agent.progress.ProgressLedger;
+import vip.mate.agent.progress.ProgressLedgerService;
+
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.concurrent.ConcurrentHashMap;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Pins the auto-record skip list in {@link ActionNode#autoRecordToolCalls}:
+ * read-only / status-query tools must NOT be written to the ledger as DONE
+ * steps. A DONE marker plus the snapshot's "已完成的步骤不要重复执行"
+ * instruction stops the agent from re-querying live state when the user
+ * repeats a question, so only mutating/task-shaped tool calls belong in
+ * the auto-recorded section.
+ */
+class ActionNodeAutoRecordSkipTest {
+
+    /** In-memory ledger double — same pattern as the service-level tests. */
+    private static final class InMemoryProgressLedgerService extends ProgressLedgerService {
+        private final Map<String, String> store = new ConcurrentHashMap<>();
+
+        InMemoryProgressLedgerService() {
+            super(null, new ObjectMapper().registerModule(new JavaTimeModule()));
+        }
+
+        @Override
+        protected String loadLedgerJson(String conversationId) {
+            return store.get(conversationId);
+        }
+
+        @Override
+        protected void saveLedgerJson(String conversationId, String json) {
+            store.put(conversationId, json);
+        }
+    }
+
+    private static ToolResponseMessage.ToolResponse resp(String name, String data) {
+        return new ToolResponseMessage.ToolResponse("id-" + name, name, data);
+    }
+
+    private static ActionNode nodeWith(ProgressLedgerService ledgerService) {
+        ActionNode node = new ActionNode(null);
+        node.setProgressLedgerService(ledgerService);
+        return node;
+    }
+
+    @Test
+    @DisplayName("read-only / status-query tools are never auto-recorded")
+    void readOnlyToolsSkipped() {
+        InMemoryProgressLedgerService ledger = new InMemoryProgressLedgerService();
+        ActionNode node = nodeWith(ledger);
+        String conv = "conv-skip-1";
+
+        node.autoRecordToolCalls(conv, List.of(
+                resp("read_file", "{\"content\":\"...\"}"),
+                resp("web_search", "results..."),
+                resp("extract_document_text", "text..."),
+                resp("extract_pdf_text", "text..."),
+                resp("extract_docx_text", "text..."),
+                resp("detect_file_type", "application/pdf"),
+                resp("getCurrentDateTime", "2026-07-31T10:33:00"),
+                resp("getCurrentDate", "2026-07-31"),
+                resp("getCurrentTime", "10:33"),
+                resp("listSubagents", "[]")));
+
+        assertTrue(ledger.load(conv).isEmpty(),
+                "no read-only tool call may produce a ledger entry");
+    }
+
+    @Test
+    @DisplayName("mutating tools are still auto-recorded alongside skipped read-only ones")
+    void mutatingToolsStillRecorded() {
+        InMemoryProgressLedgerService ledger = new InMemoryProgressLedgerService();
+        ActionNode node = nodeWith(ledger);
+        String conv = "conv-skip-2";
+
+        node.autoRecordToolCalls(conv, List.of(
+                resp("read_file", "{\"content\":\"...\"}"),
+                resp("write_file", "written"),
+                resp("mcp_home_check_room_a1b2c3", "0人")));
+
+        ProgressLedger loaded = ledger.load(conv);
+        Set<String> keys = loaded.asMap().keySet();
+        assertEquals(2, keys.size(), "exactly the non-read-only calls are recorded; got " + keys);
+        assertTrue(keys.contains(ProgressLedger.AUTO_RECORDED_PREFIX + "write_file"));
+        assertTrue(keys.contains(ProgressLedger.AUTO_RECORDED_PREFIX + "mcp_home_check_room_a1b2c3"));
+    }
+
+    @Test
+    @DisplayName("meta-tools (load_skill / enable_tool / progress_update / skill helpers) stay skipped")
+    void metaToolsStillSkipped() {
+        InMemoryProgressLedgerService ledger = new InMemoryProgressLedgerService();
+        ActionNode node = nodeWith(ledger);
+        String conv = "conv-skip-3";
+
+        node.autoRecordToolCalls(conv, List.of(
+                resp("load_skill", "loaded"),
+                resp("enable_tool", "enabled"),
+                resp("progress_update", "ok"),
+                resp("listAvailableSkills", "[]"),
+                resp("readSkillFile", "..."),
+                resp("runSkillScript", "...")));
+
+        assertTrue(ledger.load(conv).isEmpty());
+    }
+}

mateclaw-server/src/test/java/vip/mate/agent/progress/ProgressLedgerClearAutoRecordedTest.java

@@ -0,0 +1,126 @@
+package vip.mate.agent.progress;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.datatype.jsr310.JavaTimeModule;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+import vip.mate.agent.progress.ProgressLedgerService.AutoRecordEntry;
+
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.atomic.AtomicInteger;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.junit.jupiter.api.Assertions.assertDoesNotThrow;
+
+/**
+ * Pins {@link ProgressLedgerService#clearAutoRecorded} — the new-user-turn
+ * housekeeping that drops auto-recorded ({@code auto_}-prefixed) entries.
+ *
+ * <p>Why it matters: an auto-recorded DONE entry that survives into the
+ * next user turn is rendered in the snapshot together with the
+ * "已完成的步骤不要重复执行" instruction, which stops the agent from
+ * re-running a status-query tool when the user repeats a question that
+ * needs fresh data — it answers from the frozen 120-char note instead.
+ * Regular (LLM-authored) entries and pinned skill constraints must survive
+ * the clear so multi-turn task tracking keeps working.
+ */
+class ProgressLedgerClearAutoRecordedTest {
+
+    /** In-memory double — same pattern as the concurrency test. */
+    private static final class InMemoryProgressLedgerService extends ProgressLedgerService {
+        private final Map<String, String> store = new ConcurrentHashMap<>();
+        final AtomicInteger saveCount = new AtomicInteger();
+
+        InMemoryProgressLedgerService() {
+            super(null, new ObjectMapper().registerModule(new JavaTimeModule()));
+        }
+
+        @Override
+        protected String loadLedgerJson(String conversationId) {
+            return store.get(conversationId);
+        }
+
+        @Override
+        protected void saveLedgerJson(String conversationId, String json) {
+            saveCount.incrementAndGet();
+            store.put(conversationId, json);
+        }
+    }
+
+    @Test
+    @DisplayName("clearAutoRecorded drops auto_ entries but keeps regular and pinned entries")
+    void clearsOnlyAutoEntries() {
+        InMemoryProgressLedgerService service = new InMemoryProgressLedgerService();
+        String conv = "conv-clear-1";
+
+        service.upsert(conv, "step_report", "Write report", ProgressStatus.PENDING, null);
+        service.upsertPinned(conv, "pin_pdf_0", "Always render via template", null);
+        service.upsertAutoRecordedBatch(conv, List.of(
+                new AutoRecordEntry("mcp_home_check_room_a1b2c3", "check_room", "0人, 电池0%"),
+                new AutoRecordEntry("write_file", "write_file", "ok")));
+        assertEquals(4, service.load(conv).size());
+
+        service.clearAutoRecorded(conv);
+
+        ProgressLedger ledger = service.load(conv);
+        assertEquals(2, ledger.size(), "only regular + pinned should survive");
+        assertTrue(ledger.asMap().containsKey("step_report"));
+        assertTrue(ledger.pinnedEntries().containsKey("pin_pdf_0"));
+        assertFalse(ledger.asMap().keySet().stream()
+                        .anyMatch(k -> k.startsWith(ProgressLedger.AUTO_RECORDED_PREFIX)),
+                "no auto_ entry may survive a clear");
+    }
+
+    @Test
+    @DisplayName("clearAutoRecorded is a no-op (no save) when the ledger has no auto_ entries")
+    void noOpWithoutAutoEntries() {
+        InMemoryProgressLedgerService service = new InMemoryProgressLedgerService();
+        String conv = "conv-clear-2";
+
+        service.upsert(conv, "step_x", "Step X", ProgressStatus.DONE, null);
+        int savesBefore = service.saveCount.get();
+
+        service.clearAutoRecorded(conv);
+
+        assertEquals(savesBefore, service.saveCount.get(),
+                "clearing an auto-free ledger must not write to the DB");
+        assertEquals(1, service.load(conv).size());
+    }
+
+    @Test
+    @DisplayName("clearAutoRecorded tolerates null/blank conversationId and empty ledgers")
+    void toleratesMissingInput() {
+        InMemoryProgressLedgerService service = new InMemoryProgressLedgerService();
+        assertDoesNotThrow(() -> service.clearAutoRecorded(null));
+        assertDoesNotThrow(() -> service.clearAutoRecorded(""));
+        assertDoesNotThrow(() -> service.clearAutoRecorded("conv-never-seen"));
+        assertNull(((InMemoryProgressLedgerService) service).store.get("conv-never-seen"),
+                "clearing an unknown conversation must not create a row");
+    }
+
+    @Test
+    @DisplayName("auto entries re-recorded after a clear behave normally (fresh note, not the old one)")
+    void reRecordAfterClearUsesFreshResult() {
+        InMemoryProgressLedgerService service = new InMemoryProgressLedgerService();
+        String conv = "conv-clear-3";
+        String tool = "mcp_home_check_room_a1b2c3";
+
+        service.upsertAutoRecordedBatch(conv, List.of(
+                new AutoRecordEntry(tool, "check_room", "0人, 电池0%")));
+        service.clearAutoRecorded(conv);
+        // Next turn: the same tool runs again and must record the NEW result —
+        // without the clear, upsertAutoRecordedBatch skips existing keys and
+        // the note stays frozen at the first call's output.
+        service.upsertAutoRecordedBatch(conv, List.of(
+                new AutoRecordEntry(tool, "check_room", "2人, 电池87%")));
+
+        ProgressEntry entry = service.load(conv).asMap()
+                .get(ProgressLedger.AUTO_RECORDED_PREFIX + tool);
+        assertEquals("2人, 电池87%", entry.getNote());
+    }
+}