fix(llm): MiMo thinking 模式 reasoning_content 多轮对话兼容修复 (#189)

MiMo V2 系列在 thinking 模式下，assistant 消息携带 tool_calls 时必须同时包含 reasoning_content，否则提供方返回 400。

- ModelFamily 新增 MIMO_THINKING 族，detect() 添加 mimo* 匹配
- FallbackPolicy 新增 XIAOMI_MIMO（patchCrossTurn=true, patchNonToolCall=true）
- 新增 ReasoningContentCache，按 tool_call_ids 回放真实推理内容
- 缓存作用范围：所有 patchCrossTurn=true 的 thinking provider（MiMo + DeepSeek）
- NodeStreamingChatHelper 流式响应完成后写入缓存

Closes #188

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

@@ -11,6 +11,7 @@ import org.springframework.ai.chat.prompt.Prompt;
 import org.springframework.web.reactive.function.client.WebClientResponseException;
 import vip.mate.channel.web.ChatStreamTracker;
 import vip.mate.llm.chatmodel.AssistantThinkingRelay;
+import vip.mate.llm.chatmodel.ReasoningContentCache;
 
 import reactor.core.Disposable;
 
@@ -1206,6 +1207,10 @@ public class NodeStreamingChatHelper {
 
         AssistantMessage assembledMessage = buildAssistantMessageWithThinking(fullContent, fullThinking, finalToolCalls);
 
+        // Cache reasoning_content for MiMo-style providers that require it on
+        // subsequent turns.
+        cacheReasoningContent(fullThinking, finalToolCalls);
+
         recordCacheMetrics(phase, promptTok, completionTok, cacheReadTok, cacheWriteTok);
         return new StreamResult(fullContent, fullThinking, assembledMessage,
                 finalToolCalls, !finalToolCalls.isEmpty(), promptTok, completionTok,
@@ -1235,6 +1240,10 @@ public class NodeStreamingChatHelper {
 
         AssistantMessage assembledMessage = buildAssistantMessageWithThinking(fullContent, fullThinking, finalToolCalls);
 
+        // Cache reasoning_content for MiMo-style providers that require it on
+        // subsequent turns. The cache replays real values instead of empty strings.
+        cacheReasoningContent(fullThinking, finalToolCalls);
+
         recordCacheMetrics(phase, promptTok, completionTok, cacheReadTok, cacheWriteTok);
         return new StreamResult(fullContent, fullThinking, assembledMessage,
                 finalToolCalls, !finalToolCalls.isEmpty(), promptTok, completionTok,
@@ -1268,6 +1277,24 @@ public class NodeStreamingChatHelper {
         return builder.build();
     }
 
+    /**
+     * Store reasoning content in the cache for cross-turn replay.
+     * Only caches when there are tool calls (MiMo requires reasoning_content
+     * specifically on assistant messages with tool_calls).
+     */
+    private static void cacheReasoningContent(String fullThinking,
+                                               List<AssistantMessage.ToolCall> toolCalls) {
+        if (fullThinking == null || fullThinking.isBlank()) return;
+        if (toolCalls == null || toolCalls.isEmpty()) return;
+        List<String> ids = toolCalls.stream()
+                .map(AssistantMessage.ToolCall::id)
+                .filter(id -> id != null && !id.isEmpty())
+                .toList();
+        if (!ids.isEmpty()) {
+            ReasoningContentCache.store(ids, fullThinking);
+        }
+    }
+
     /**
      * Record token / cache usage to the optional metrics aggregator.
      * Called only from successful assembly paths ({@link #assembleResult}

mateclaw-server/src/main/java/vip/mate/llm/chatmodel/OpenAiRequestRewriter.java

@@ -137,12 +137,17 @@ final class OpenAiRequestRewriter {
             if (next != null && !next.isEmpty()) {
                 injected = next;
             } else {
-                injected = policy.emptyFallback;
-                if (injected == null && policy.warnOnMissingReal) {
-                    log.warn("[patchReasoningContent] provider={} requires real reasoning_content "
-                                    + "but relay has no value for assistant message at index {}; "
-                                    + "leaving null so provider returns explicit error.",
-                            providerIdOrUnknown(provider), i);
+                // For cross-turn messages, try the reasoning content cache
+                // (real values from prior responses) before falling back to empty.
+                injected = resolveCrossTurnReasoning(msg, i <= lastUserIdx);
+                if (injected == null) {
+                    injected = policy.emptyFallback;
+                    if (injected == null && policy.warnOnMissingReal) {
+                        log.warn("[patchReasoningContent] provider={} requires real reasoning_content "
+                                        + "but relay has no value for assistant message at index {}; "
+                                        + "leaving null so provider returns explicit error.",
+                                providerIdOrUnknown(provider), i);
+                    }
                 }
             }
             if (injected == null && msg.reasoningContent() == null) {
@@ -225,10 +230,11 @@ final class OpenAiRequestRewriter {
      * OpenAI-compatible gateway) might still require the patch.
      */
     private enum FallbackPolicy {
-        DEEPSEEK(" ",  false, true,  true),
-        KIMI    (" ",  false, false, false),
-        OPENAI  (" ",  false, false, false),
-        DEFAULT (" ",  false, false, false);
+        DEEPSEEK    (" ",  false, true,  true),
+        KIMI        (" ",  false, false, false),
+        OPENAI      (" ",  false, false, false),
+        XIAOMI_MIMO (" ",  false, true,  true),
+        DEFAULT     (" ",  false, false, false);
 
         final String emptyFallback;
         final boolean warnOnMissingReal;
@@ -253,6 +259,7 @@ final class OpenAiRequestRewriter {
                 case "deepseek" -> DEEPSEEK;
                 case "kimi-cn", "kimi-intl", "kimi-code" -> KIMI;
                 case "openai", "azure-openai" -> OPENAI;
+                case "xiaomi-mimo" -> XIAOMI_MIMO;
                 default -> DEFAULT;
             };
         }
@@ -306,6 +313,22 @@ final class OpenAiRequestRewriter {
         return family.isThinking();
     }
 
+    /**
+     * Look up cached reasoning content for cross-turn assistant messages.
+     * Returns the cached value, or {@code null} if no cache hit (caller falls
+     * back to the policy's empty fallback).
+     */
+    private static String resolveCrossTurnReasoning(
+            OpenAiApi.ChatCompletionMessage msg, boolean isCrossTurn) {
+        if (!isCrossTurn) return null;
+        if (msg.toolCalls() == null || msg.toolCalls().isEmpty()) return null;
+        List<String> ids = msg.toolCalls().stream()
+                .map(tc -> tc.id())
+                .filter(id -> id != null && !id.isEmpty())
+                .toList();
+        return ReasoningContentCache.get(ids);
+    }
+
     // ==================== reasoning_effort sanitizing ====================
 
     /**

mateclaw-server/src/main/java/vip/mate/llm/chatmodel/ReasoningContentCache.java

@@ -0,0 +1,116 @@
+package vip.mate.llm.chatmodel;
+
+import lombok.extern.slf4j.Slf4j;
+
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
+
+/**
+ * Static singleton cache for MiMo-style {@code reasoning_content} replay.
+ *
+ * <p>MiMo (and similar providers) require {@code reasoning_content} on assistant
+ * messages that carry {@code tool_calls}. When the conversation spans multiple
+ * turns, the cache replays the <em>real</em> reasoning content from prior
+ * responses instead of injecting empty strings, preserving model context.
+ *
+ * <h2>Key design</h2>
+ * <ul>
+ *   <li>Key: sorted, concatenated tool_call IDs from the assistant message.
+ *       Tool call IDs are unique per response, so this key naturally
+ *       disambiguates across turns.</li>
+ *   <li>TTL: 24 hours (configurable). Entries older than TTL are lazily evicted
+ *       on access and periodically during {@link #store}.</li>
+ *   <li>Max entries: 10,000. Oldest entries evicted when exceeded.</li>
+ * </ul>
+ *
+ * <h2>Usage</h2>
+ * <ol>
+ *   <li><b>Store</b>: after a streaming response completes, call
+ *       {@link #store} with the tool_call IDs and reasoning content.</li>
+ *   <li><b>Retrieve</b>: during request patching, call {@link #get} for
+ *       cross-turn assistant messages to fill in cached reasoning.</li>
+ * </ol>
+ *
+ * @author MateClaw Team
+ */
+@Slf4j
+public final class ReasoningContentCache {
+
+    private static final long DEFAULT_MAX_AGE_MS = 24 * 60 * 60 * 1000L; // 24 hours
+    private static final int DEFAULT_MAX_ENTRIES = 10_000;
+    private static final long EVICT_INTERVAL_MS = 5 * 60 * 1000L; // 5 minutes
+
+    private static final ConcurrentHashMap<String, Entry> MAP = new ConcurrentHashMap<>();
+    private static volatile long lastEvictMs = System.currentTimeMillis();
+
+    private ReasoningContentCache() {}
+
+    /**
+     * Cache reasoning content for a set of tool_call IDs.
+     *
+     * @param toolCallIds     tool call IDs from the assistant message (must not be null/empty)
+     * @param reasoningContent the real reasoning content to cache (must not be blank)
+     */
+    public static void store(List<String> toolCallIds, String reasoningContent) {
+        if (toolCallIds == null || toolCallIds.isEmpty()) return;
+        if (reasoningContent == null || reasoningContent.isBlank()) return;
+
+        String key = makeKey(toolCallIds);
+        MAP.put(key, new Entry(reasoningContent, System.currentTimeMillis()));
+
+        maybeEvict();
+    }
+
+    /**
+     * Retrieve cached reasoning content for the given tool_call IDs.
+     *
+     * @return cached reasoning content, or {@code null} if not found or expired
+     */
+    public static String get(List<String> toolCallIds) {
+        if (toolCallIds == null || toolCallIds.isEmpty()) return null;
+
+        String key = makeKey(toolCallIds);
+        Entry entry = MAP.get(key);
+        if (entry == null) return null;
+
+        if (System.currentTimeMillis() - entry.storedAtMs > DEFAULT_MAX_AGE_MS) {
+            MAP.remove(key);
+            return null;
+        }
+        return entry.reasoningContent;
+    }
+
+    /** Clear all cached entries. */
+    public static void clear() {
+        MAP.clear();
+    }
+
+    /** Current cache size (for diagnostics). */
+    public static int size() {
+        return MAP.size();
+    }
+
+    private static String makeKey(List<String> toolCallIds) {
+        return String.join("|", toolCallIds.stream().sorted().toList());
+    }
+
+    private static void maybeEvict() {
+        long now = System.currentTimeMillis();
+        if (now - lastEvictMs < EVICT_INTERVAL_MS) return;
+        lastEvictMs = now;
+
+        // Remove expired entries
+        MAP.entrySet().removeIf(e -> now - e.getValue().storedAtMs > DEFAULT_MAX_AGE_MS);
+
+        // Remove oldest if over limit
+        if (MAP.size() > DEFAULT_MAX_ENTRIES) {
+            MAP.entrySet().stream()
+                    .sorted((a, b) -> Long.compare(a.getValue().storedAtMs, b.getValue().storedAtMs))
+                    .limit(MAP.size() - DEFAULT_MAX_ENTRIES)
+                    .forEach(e -> MAP.remove(e.getKey()));
+        }
+    }
+
+    private record Entry(String reasoningContent, long storedAtMs) {}
+}

mateclaw-server/src/main/java/vip/mate/llm/model/ModelFamily.java

@@ -58,6 +58,15 @@ public enum ModelFamily {
      */
     DEEPSEEK_V4_REASONING(false, false, true, false, false, true),
 
+    /**
+     * Xiaomi MiMo thinking 模型：MiMo-VL-*、mimo-* 系列。
+     * <p>
+     * MiMo 的 thinking 模式与 DeepSeek 类似：响应返回 {@code reasoning_content}，
+     * 后续多轮请求必须将 {@code reasoning_content} 传回，否则 API 返回 400 错误。
+     * 约束：保留 max_tokens；不支持 reasoning_effort；temperature/topP 用配置值。
+     */
+    MIMO_THINKING(false, false, false, false, false, true),
+
     /**
      * 通用 thinking 模型（名称含 "thinking" 或 "reasoner" 但不匹配上述族）：
      * 如 qwen3-235b-a22b-thinking-2507
@@ -160,6 +169,11 @@ public enum ModelFamily {
             return DEEPSEEK_REASONER;
         }
 
+        // Xiaomi MiMo thinking 族：mimo-* / MiMo-VL-* 系列
+        if (normalized.startsWith("mimo")) {
+            return MIMO_THINKING;
+        }
+
         // 通用 thinking 族：名称含 thinking / reasoner 关键词
         if (normalized.contains("thinking") || normalized.contains("reasoner")) {
             return GENERIC_THINKING;

mateclaw-server/src/test/java/vip/mate/llm/chatmodel/ReasoningContentCacheTest.java

@@ -0,0 +1,69 @@
+package vip.mate.llm.chatmodel;
+
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import java.util.List;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+class ReasoningContentCacheTest {
+
+    @AfterEach
+    void cleanup() {
+        ReasoningContentCache.clear();
+    }
+
+    @Test
+    @DisplayName("Store and retrieve reasoning content by tool_call IDs")
+    void storeAndGet() {
+        List<String> ids = List.of("call_1", "call_2");
+        ReasoningContentCache.store(ids, "thinking content here");
+
+        assertEquals("thinking content here", ReasoningContentCache.get(ids));
+    }
+
+    @Test
+    @DisplayName("Cache key is order-independent (sorted tool_call IDs)")
+    void orderIndependent() {
+        ReasoningContentCache.store(List.of("call_b", "call_a"), "content");
+
+        assertEquals("content", ReasoningContentCache.get(List.of("call_a", "call_b")));
+    }
+
+    @Test
+    @DisplayName("Miss returns null")
+    void cacheMiss() {
+        assertNull(ReasoningContentCache.get(List.of("nonexistent")));
+    }
+
+    @Test
+    @DisplayName("Empty/null tool_call IDs are no-ops")
+    void emptyIds() {
+        ReasoningContentCache.store(List.of(), "content");
+        ReasoningContentCache.store(null, "content");
+        assertEquals(0, ReasoningContentCache.size());
+    }
+
+    @Test
+    @DisplayName("Blank/null reasoning content is not cached")
+    void blankContent() {
+        ReasoningContentCache.store(List.of("call_1"), "");
+        ReasoningContentCache.store(List.of("call_1"), "  ");
+        ReasoningContentCache.store(List.of("call_1"), null);
+        assertEquals(0, ReasoningContentCache.size());
+    }
+
+    @Test
+    @DisplayName("Clear removes all entries")
+    void clearAll() {
+        ReasoningContentCache.store(List.of("call_1"), "content1");
+        ReasoningContentCache.store(List.of("call_2"), "content2");
+        assertEquals(2, ReasoningContentCache.size());
+
+        ReasoningContentCache.clear();
+        assertEquals(0, ReasoningContentCache.size());
+        assertNull(ReasoningContentCache.get(List.of("call_1")));
+    }
+}

mateclaw-server/src/test/java/vip/mate/llm/model/ModelFamilyTest.java

@@ -65,4 +65,15 @@ class ModelFamilyTest {
         assertEquals(ModelFamily.STANDARD, ModelFamily.detect(""));
         assertEquals(ModelFamily.STANDARD, ModelFamily.detect("   "));
     }
+
+    @Test
+    @DisplayName("Xiaomi MiMo models → MIMO_THINKING (reasoning_content relay required)")
+    void mimo_thinking() {
+        assertEquals(ModelFamily.MIMO_THINKING, ModelFamily.detect("mimo-v2-flash"));
+        assertEquals(ModelFamily.MIMO_THINKING, ModelFamily.detect("MiMo-VL-7B-RL"));
+        assertTrue(ModelFamily.MIMO_THINKING.isThinking(),
+                "Mimo must be flagged as thinking so reasoning_content is patched");
+        assertFalse(ModelFamily.MIMO_THINKING.supportsReasoningEffort(),
+                "Mimo does not accept the reasoning_effort parameter");
+    }
 }