feat(agent): 小上下文降级档——紧凑/底线档收紧注入、压缩触发比自适应、prefix 分块统计与预超限快速失败

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

@@ -3,6 +3,7 @@ package vip.mate.agent.context;
 import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.ai.chat.messages.AssistantMessage;
+import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.ai.chat.messages.Message;
 import org.springframework.ai.chat.messages.SystemMessage;
 import org.springframework.ai.chat.messages.ToolResponseMessage;
@@ -118,6 +119,18 @@ public class ConversationWindowManager {
     private final MemoryManager memoryManager;
     private final ConversationService conversationService;
 
+    /**
+     * Optional — adaptive compaction trigger for small context windows.
+     * Setter-injected so the many direct test constructions keep the plain
+     * configured ratio (null → previous behavior).
+     */
+    private PrefixBudgetPlanner prefixBudgetPlanner;
+
+    @Autowired(required = false)
+    public void setPrefixBudgetPlanner(PrefixBudgetPlanner prefixBudgetPlanner) {
+        this.prefixBudgetPlanner = prefixBudgetPlanner;
+    }
+
     /**
      * Optional spill store, injected via setter so unit tests and the two
      * existing 3-arg constructor callers in tests stay source-compatible.
@@ -251,7 +264,12 @@ public class ConversationWindowManager {
 
         int effectiveMax = (maxInputTokens != null && maxInputTokens > 0)
                 ? maxInputTokens : properties.getDefaultMaxInputTokens();
-        int triggerThreshold = (int) (effectiveMax * properties.getCompactTriggerRatio());
+        // Small windows compact later (higher trigger ratio): summarizing at
+        // 75% of an 8k window throws away room it cannot afford to lose.
+        double triggerRatio = prefixBudgetPlanner != null
+                ? prefixBudgetPlanner.compactTriggerRatioFor(effectiveMax, properties.getCompactTriggerRatio())
+                : properties.getCompactTriggerRatio();
+        int triggerThreshold = (int) (effectiveMax * triggerRatio);
 
         int systemTokens = TokenEstimator.estimateTokens(systemPrompt);
         int currentMsgTokens = TokenEstimator.estimateTokens(currentUserMessage) + TokenEstimator.PER_MESSAGE_OVERHEAD;

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

@@ -30,6 +30,9 @@ import vip.mate.config.PrefixBudgetProperties;
 @EnableConfigurationProperties(PrefixBudgetProperties.class)
 public class PrefixBudgetPlanner {
 
+    /** COMPACT-profile ceiling for the wiki relevance injection (~one page). */
+    static final int COMPACT_WIKI_TOKEN_CAP = 2000;
+
     private final PrefixBudgetProperties properties;
     private final ConversationWindowProperties windowProperties;
 
@@ -66,10 +69,21 @@ public class PrefixBudgetPlanner {
             sum = 1.0;
         }
 
+        // Profile-specific wiki clamps: knowledge-base reference pages are the
+        // most dispensable block on a small window — the wiki tools stay
+        // callable, only the automatic pre-injection shrinks. COMPACT caps it
+        // at roughly one page; MINIMAL disables it outright.
+        int wikiTokens = (int) (injectionBudget * shares.getWiki() / sum);
+        wikiTokens = switch (profile) {
+            case NORMAL -> wikiTokens;
+            case COMPACT -> Math.min(wikiTokens, COMPACT_WIKI_TOKEN_CAP);
+            case MINIMAL -> 0;
+        };
+
         PrefixBudgetPlan plan = new PrefixBudgetPlan(
                 true, effectiveMax, profile, injectionBudget,
                 (int) (injectionBudget * shares.getMemory() / sum),
-                (int) (injectionBudget * shares.getWiki() / sum),
+                wikiTokens,
                 (int) (injectionBudget * shares.getSkill() / sum),
                 (int) (injectionBudget * shares.getExtensionCatalog() / sum),
                 (int) (injectionBudget * shares.getLedger() / sum),

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

@@ -749,6 +749,27 @@ public class ReasoningNode implements NodeAction {
 
         Prompt prompt = new Prompt(promptMessages, options);
 
+        // Prefix accounting: how much of the window the never-trimmed prefix
+        // (system prompt + runtime context + wiki + skill catalog + ledger)
+        // and the advertised tool schemas consume. Logged on the turn's first
+        // call so a small-window overflow is diagnosable per block instead of
+        // surfacing as an opaque provider 400.
+        int prefixEstimateTokens = TokenEstimator.estimateTokens(nonHistoryPrefix);
+        int toolSchemaEstimateTokens = TokenEstimator.estimateToolsTokens(activeCallbacks);
+        if (accessor.llmCallCount() == 0) {
+            log.info("[ReasoningNode] Prefix accounting conv={}: window={} tokens, prefix={} "
+                            + "(system+context+wiki+skills+ledger), toolSchemas={}, history={}",
+                    conversationId, loopContextWindowTokens(), prefixEstimateTokens,
+                    toolSchemaEstimateTokens, TokenEstimator.estimateTokens(messages));
+        }
+        // The prefix cannot be compacted (history compaction is the only lever),
+        // so a prefix that alone exceeds the window makes the request doomed —
+        // fail fast with the same PROMPT_TOO_LONG shape a provider rejection
+        // would produce instead of sending it. Gated on budgeting being active
+        // (an estimation false-positive must not block requests otherwise).
+        boolean prefixOverflow = prefixBudgetPlan != null && prefixBudgetPlan.enabled()
+                && prefixEstimateTokens + toolSchemaEstimateTokens > prefixBudgetPlan.effectiveMaxTokens();
+
         // ======= LLM 调用区域 =======
         // nextLlmCallCount 在首次 streamCall 之前计算。
         // 所有退出路径（正常、stopped、fatal error、CancellationException）都必须写回此值。
@@ -778,7 +799,19 @@ public class ReasoningNode implements NodeAction {
 
         NodeStreamingChatHelper.StreamResult result;
         try {
-            result = streamingHelper.streamCall(chatModel, prompt, conversationId, "reasoning");
+            if (prefixOverflow) {
+                String overflowMessage = "Prompt 前缀估算 " + (prefixEstimateTokens + toolSchemaEstimateTokens)
+                        + " tokens(注入块 " + prefixEstimateTokens + " + 工具 schema " + toolSchemaEstimateTokens
+                        + ")已超过模型上下文窗口 " + prefixBudgetPlan.effectiveMaxTokens()
+                        + " tokens,历史压缩无法解决——请精简 Agent 身份 prompt、减少绑定工具/技能,"
+                        + "或换用更大窗口的模型";
+                log.error("[ReasoningNode] {}", overflowMessage);
+                result = new NodeStreamingChatHelper.StreamResult(null, null, null, List.of(), false,
+                        0, 0, false, overflowMessage,
+                        NodeStreamingChatHelper.ErrorType.PROMPT_TOO_LONG, false, 0, 0, 0);
+            } else {
+                result = streamingHelper.streamCall(chatModel, prompt, conversationId, "reasoning");
+            }
 
             // PTL 处理：结构化压缩后重试。复用 nonHistoryPrefix 保证重试
             // Prompt 仍带 wiki / runtime context；早期的 tail-only 路径会把

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

@@ -44,11 +44,35 @@ class PrefixBudgetPlannerTest {
     }
 
     @Test
-    @DisplayName("4k window → MINIMAL profile")
+    @DisplayName("4k window → MINIMAL profile, wiki injection disabled outright")
     void minimalProfile() {
         PrefixBudgetPlan plan = planner.plan(4096, 500, 500);
         assertEquals(PrefixBudgetPlan.Profile.MINIMAL, plan.profile());
         assertEquals(Math.max(0, (int) (4096 * 0.15) - 1000), plan.injectionBudgetTokens());
+        assertEquals(0, plan.wikiTokens());
+    }
+
+    @Test
+    @DisplayName("COMPACT profile caps the wiki injection at roughly one page")
+    void compactCapsWiki() {
+        properties.getShares().setWiki(1.0);
+        properties.getShares().setMemory(0.0);
+        properties.getShares().setSkill(0.0);
+        properties.getShares().setExtensionCatalog(0.0);
+        properties.getShares().setLedger(0.0);
+        PrefixBudgetPlan plan = planner.plan(20000, 0, 0);
+        assertEquals(PrefixBudgetPlan.Profile.COMPACT, plan.profile());
+        assertTrue(plan.injectionBudgetTokens() > PrefixBudgetPlanner.COMPACT_WIKI_TOKEN_CAP);
+        assertEquals(PrefixBudgetPlanner.COMPACT_WIKI_TOKEN_CAP, plan.wikiTokens());
+    }
+
+    @Test
+    @DisplayName("plan carries an independent tool-schema budget from toolSchemaRatio")
+    void toolSchemaBudget() {
+        PrefixBudgetPlan plan = planner.plan(16384, 0, 0);
+        assertEquals((int) (16384 * 0.25), plan.toolSchemaBudgetTokens());
+        properties.setEnabled(false);
+        assertEquals(Integer.MAX_VALUE, planner.plan(16384, 0, 0).toolSchemaBudgetTokens());
     }
 
     @Test