diff --git a/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java b/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java index 0af112c7..2bcddb83 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java @@ -303,7 +303,7 @@ public class AgentGraphBuilder { String reasoningEffort = resolveReasoningEffortForModel(runtimeModel); CompiledGraph compiledGraph = buildReActGraph(toolSet, chatModel, maxIter, reasoningEffort, runtimeModel, agentId); return new StateGraphReActAgent(chatClient, conversationService, compiledGraph, - chatModel, conversationWindowManager); + chatModel, conversationWindowManager, toolSet); } StateGraphPlanExecuteAgent buildPlanExecuteAgent(AgentToolSet toolSet, ModelConfigEntity runtimeModel, int maxIter) { @@ -317,7 +317,7 @@ public class AgentGraphBuilder { String reasoningEffort = resolveReasoningEffortForModel(runtimeModel); CompiledGraph graph = buildPlanExecuteGraph(toolSet, chatModel, maxIter, reasoningEffort, runtimeModel, agentId); return new StateGraphPlanExecuteAgent(chatClient, conversationService, graph, planningService, - chatModel, conversationWindowManager); + chatModel, conversationWindowManager, toolSet); } CompiledGraph buildPlanExecuteGraph(AgentToolSet toolSet, ChatModel chatModel, int maxIterations, String reasoningEffort) { diff --git a/mateclaw-server/src/main/java/vip/mate/agent/context/ConversationWindowManager.java b/mateclaw-server/src/main/java/vip/mate/agent/context/ConversationWindowManager.java index 7b6782fc..e9214773 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/context/ConversationWindowManager.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/context/ConversationWindowManager.java @@ -11,6 +11,7 @@ import org.springframework.ai.chat.model.ChatModel; import org.springframework.ai.chat.model.ChatResponse; import org.springframework.ai.chat.prompt.ChatOptions; import org.springframework.ai.chat.prompt.Prompt; +import org.springframework.ai.tool.ToolCallback; import com.alibaba.cloud.ai.dashscope.chat.DashScopeChatOptions; import org.springframework.stereotype.Component; import vip.mate.agent.prompt.PromptLoader; @@ -116,6 +117,23 @@ public class ConversationWindowManager { String currentUserMessage, Integer maxInputTokens, ChatModel chatModel, String conversationId, Long agentId) { + return fitToWindow(messages, systemPrompt, currentUserMessage, + maxInputTokens, chatModel, conversationId, agentId, null); + } + + /** + * Same as the 7-arg overload but additionally accounts for the tool + * definitions sent on every LLM call. Without {@code toolCallbacks}, + * the budget calculation underestimates the actual request size by the + * full size of the tools schema (often several thousand tokens for + * agents bound to multiple MCP servers), making compression fire too + * late and producing HTTP 400 once the request hits the model. + */ + public List fitToWindow(List messages, String systemPrompt, + String currentUserMessage, + Integer maxInputTokens, ChatModel chatModel, + String conversationId, Long agentId, + java.util.Collection toolCallbacks) { if (messages == null || messages.isEmpty()) { return messages; } @@ -127,20 +145,21 @@ public class ConversationWindowManager { int systemTokens = TokenEstimator.estimateTokens(systemPrompt); int currentMsgTokens = TokenEstimator.estimateTokens(currentUserMessage) + TokenEstimator.PER_MESSAGE_OVERHEAD; int historyTokens = TokenEstimator.estimateTokens(messages); - int totalTokens = systemTokens + currentMsgTokens + historyTokens; + int toolsTokens = TokenEstimator.estimateToolsTokens(toolCallbacks); + int totalTokens = systemTokens + currentMsgTokens + historyTokens + toolsTokens; if (totalTokens <= triggerThreshold) { return messages; } - log.info("[ConversationWindow] 超阈值: {} tokens (system={}, current={}, history={}) > {} 触发阈值 (max={}), conv={}", - totalTokens, systemTokens, currentMsgTokens, historyTokens, + log.info("[ConversationWindow] 超阈值: {} tokens (system={}, current={}, history={}, tools={}) > {} 触发阈值 (max={}), conv={}", + totalTokens, systemTokens, currentMsgTokens, historyTokens, toolsTokens, triggerThreshold, effectiveMax, conversationId); evictExpiredEntries(); - // 可用于历史的 token 预算 = max - system - currentMsg - 安全余量 - int reservedTokens = systemTokens + currentMsgTokens + (int) (effectiveMax * 0.05); + // 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%。 // 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 diff --git a/mateclaw-server/src/main/java/vip/mate/agent/context/TokenEstimator.java b/mateclaw-server/src/main/java/vip/mate/agent/context/TokenEstimator.java index 0f14132e..cb974fae 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/context/TokenEstimator.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/context/TokenEstimator.java @@ -1,7 +1,10 @@ package vip.mate.agent.context; import org.springframework.ai.chat.messages.Message; +import org.springframework.ai.tool.ToolCallback; +import org.springframework.ai.tool.definition.ToolDefinition; +import java.util.Collection; import java.util.List; /** @@ -21,6 +24,13 @@ public final class TokenEstimator { /** 每条消息的固定开销 token（role 标记、分隔符等） */ static final int PER_MESSAGE_OVERHEAD = 4; + /** + * Per-tool wrapper overhead: function/type:object boilerplate, name and + * description framing, parameters key, and JSON braces around the schema. + * Conservative — slightly overestimates so budget guards don't underrun. + */ + static final int PER_TOOL_OVERHEAD = 12; + private TokenEstimator() { } @@ -71,6 +81,38 @@ public final class TokenEstimator { .sum(); } + /** + * Estimate the token cost of the tool definitions sent on every LLM call + * (name + description + JSON inputSchema, plus per-tool wrapper overhead). + *

+ * A heavily-bound agent (multiple MCP servers, many built-ins) can carry + * several thousand tokens of tool schema on every request — leaving them + * out of the context-window budget makes compression decisions fire too + * late and on small models triggers HTTP 400 once the request actually + * goes out. + */ + public static int estimateToolsTokens(Collection callbacks) { + if (callbacks == null || callbacks.isEmpty()) { + return 0; + } + int total = 0; + for (ToolCallback cb : callbacks) { + if (cb == null) continue; + ToolDefinition def; + try { + def = cb.getToolDefinition(); + } catch (Exception e) { + continue; + } + if (def == null) continue; + total += estimateTokens(def.name()) + + estimateTokens(def.description()) + + estimateTokens(def.inputSchema()) + + PER_TOOL_OVERHEAD; + } + return total; + } + /** * 判断是否为 CJK 字符（中日韩统一表意文字 + 常用标点） */ diff --git a/mateclaw-server/src/main/java/vip/mate/agent/graph/StateGraphReActAgent.java b/mateclaw-server/src/main/java/vip/mate/agent/graph/StateGraphReActAgent.java index 64c65d4c..8afd7ae3 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/graph/StateGraphReActAgent.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/graph/StateGraphReActAgent.java @@ -53,15 +53,32 @@ public class StateGraphReActAgent extends BaseAgent implements StructuredStreamC private final CompiledGraph compiledGraph; private final org.springframework.ai.chat.model.ChatModel chatModel; private final ConversationWindowManager conversationWindowManager; + /** + * Held only so {@link #buildInitialState} can include the tools schema in + * the context-window budget — those bytes ride along on every LLM call + * and were previously ignored, making compression decisions fire late. + * Nullable for the legacy 5-arg constructor used by older tests. + */ + private final vip.mate.agent.AgentToolSet toolSet; public StateGraphReActAgent(ChatClient chatClient, ConversationService conversationService, CompiledGraph compiledGraph, org.springframework.ai.chat.model.ChatModel chatModel, ConversationWindowManager conversationWindowManager) { + this(chatClient, conversationService, compiledGraph, chatModel, + conversationWindowManager, null); + } + + public StateGraphReActAgent(ChatClient chatClient, ConversationService conversationService, + CompiledGraph compiledGraph, + org.springframework.ai.chat.model.ChatModel chatModel, + ConversationWindowManager conversationWindowManager, + vip.mate.agent.AgentToolSet toolSet) { super(chatClient, conversationService); this.compiledGraph = compiledGraph; this.chatModel = chatModel; this.conversationWindowManager = conversationWindowManager; + this.toolSet = toolSet; } @Override @@ -372,7 +389,8 @@ public class StateGraphReActAgent extends BaseAgent implements StructuredStreamC maxInputTokens, chatModel, conversationId, - parsedAgentId); + parsedAgentId, + toolSet != null ? toolSet.callbacks() : null); } List messages = new ArrayList<>(historyMessages); diff --git a/mateclaw-server/src/main/java/vip/mate/agent/graph/plan/StateGraphPlanExecuteAgent.java b/mateclaw-server/src/main/java/vip/mate/agent/graph/plan/StateGraphPlanExecuteAgent.java index 64a25303..39cd4ecf 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/graph/plan/StateGraphPlanExecuteAgent.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/graph/plan/StateGraphPlanExecuteAgent.java @@ -45,16 +45,28 @@ public class StateGraphPlanExecuteAgent extends BaseAgent implements StructuredS private final PlanningService planningService; private final org.springframework.ai.chat.model.ChatModel chatModel; private final ConversationWindowManager conversationWindowManager; + /** Held only so context-window budget includes the tools schema. Nullable for legacy constructor. */ + private final vip.mate.agent.AgentToolSet toolSet; public StateGraphPlanExecuteAgent(ChatClient chatClient, ConversationService conversationService, CompiledGraph compiledGraph, PlanningService planningService, org.springframework.ai.chat.model.ChatModel chatModel, ConversationWindowManager conversationWindowManager) { + this(chatClient, conversationService, compiledGraph, planningService, + chatModel, conversationWindowManager, null); + } + + public StateGraphPlanExecuteAgent(ChatClient chatClient, ConversationService conversationService, + CompiledGraph compiledGraph, PlanningService planningService, + org.springframework.ai.chat.model.ChatModel chatModel, + ConversationWindowManager conversationWindowManager, + vip.mate.agent.AgentToolSet toolSet) { super(chatClient, conversationService); this.compiledGraph = compiledGraph; this.planningService = planningService; this.chatModel = chatModel; this.conversationWindowManager = conversationWindowManager; + this.toolSet = toolSet; } @Override @@ -254,7 +266,8 @@ public class StateGraphPlanExecuteAgent extends BaseAgent implements StructuredS maxInputTokens, chatModel, conversationId, - parsedAgentId); + parsedAgentId, + toolSet != null ? toolSet.callbacks() : null); } List messages = new ArrayList<>(historyMessages);