@mingxy/cerebro 1.15.14 → 1.16.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mingxy/cerebro",
-  "version": "1.15.14",
+  "version": "1.16.0",
   "description": "Cerebro persistent memory plugin for OpenCode — auto-recall, auto-capture, 9 memory tools with clustering, project-scoped memory isolation",
   "type": "module",
   "main": "src/index.ts",

package/schema.json

@@ -144,13 +144,6 @@
           "description": "LLM refinement strategy: strict (high only), balanced (high+medium), loose (keep all)",
           "enum": ["strict", "balanced", "loose"],
           "default": "balanced"
-        },
-        "refineMediumChars": {
-          "type": "number",
-          "description": "Character limit for medium-relevance content truncation",
-          "default": 200,
-          "minimum": 50,
-          "maximum": 2000
         }
       },
       "additionalProperties": false

package/src/client.ts

@@ -363,7 +363,6 @@ export class CerebroClient {
       phase2_multiplier?: number;
       llm_max_eval?: number;
       refine_strategy?: string;
-      refine_medium_chars?: number;
       skip_llm_gate?: boolean;
     },
     projectPath?: string,

package/src/config.ts

@@ -29,7 +29,6 @@ export interface OmemPluginConfig {
     phase2Multiplier: number;
     llmMaxEval: number;
     refineStrategy: "strict" | "balanced" | "loose";
-    refineMediumChars: number;
   };
   logging: {
     logEnabled: boolean;
@@ -73,7 +72,6 @@ const DEFAULTS: OmemPluginConfig = {
     phase2Multiplier: 2,
     llmMaxEval: 15,
     refineStrategy: "balanced",
-    refineMediumChars: 200,
   },
   logging: {
     logEnabled: true,
@@ -140,7 +138,6 @@ function migrateFlatToNested(flat: FlatConfig): OmemPluginConfig {
       phase2Multiplier: DEFAULTS.recall.phase2Multiplier,
       llmMaxEval: DEFAULTS.recall.llmMaxEval,
       refineStrategy: DEFAULTS.recall.refineStrategy,
-      refineMediumChars: DEFAULTS.recall.refineMediumChars,
     },
     logging: {
       logEnabled: flat.logEnabled ?? DEFAULTS.logging.logEnabled,

package/src/hooks.ts

@@ -243,14 +243,30 @@ const FETCH_POLICY = [
   "</cerebro-fetch-policy>",
 ].join("\n");
 
-function buildContextBlock(results: SearchResult[], maxContentLength: number = 500): string {
+/**
+ * Score-weighted budget allocation: high-score memories get more chars.
+ * Falls back to uniform distribution when totalScore === 0 or all scores equal.
+ */
+function buildContextBlock(
+  results: SearchResult[],
+  budget: number,
+  maxContentLength: number = 500,
+  minItemChars: number = MIN_ITEM_CONTENT_CHARS,
+): string {
   if (results.length === 0) return "";
 
+  const totalScore = results.reduce((sum, r) => sum + r.score, 0);
+
   const grouped = categorize(results);
   const sections: string[] = [];
 
   for (const [label, items] of grouped) {
-    const lines = items.map((r) => formatMemoryLine(r, maxContentLength));
+    const lines = items.map((r) => {
+      const itemMaxLen = totalScore > 0
+        ? Math.min(maxContentLength, Math.max(minItemChars, Math.floor((r.score / totalScore) * budget)))
+        : Math.min(maxContentLength, Math.max(minItemChars, Math.floor(budget / results.length)));
+      return formatMemoryLine(r, itemMaxLen);
+    });
     sections.push(`[${label}]\n${lines.join("\n")}`);
   }
 
@@ -262,13 +278,23 @@ function buildContextBlock(results: SearchResult[], maxContentLength: number = 5
   ].join("\n");
 }
 
-function buildClusteredContextBlock(clustered: import("./client.js").ClusteredRecallResult, maxContentLength: number = 500): string {
+function buildClusteredContextBlock(
+  clustered: import("./client.js").ClusteredRecallResult,
+  budget: number,
+  maxContentLength: number = 500,
+  minItemChars: number = MIN_ITEM_CONTENT_CHARS,
+): string {
   const sections: string[] = [];
 
   if (clustered.cluster_summaries.length > 0) {
+    const totalClusterScore = clustered.cluster_summaries.reduce((sum, cs) => sum + cs.relevance_score, 0);
+
     sections.push("## 📋 主题簇（聚合记忆）");
     for (const cs of clustered.cluster_summaries) {
       const scoreIndicator = cs.relevance_score >= 0.8 ? "★★★" : cs.relevance_score >= 0.6 ? "★★" : "★";
+      const clusterMaxLen = totalClusterScore > 0
+        ? Math.min(maxContentLength, Math.max(minItemChars, Math.floor((cs.relevance_score / totalClusterScore) * budget)))
+        : Math.min(maxContentLength, Math.max(minItemChars, Math.floor(budget / clustered.cluster_summaries.length)));
       sections.push(`\n### ${cs.title} (整合自${cs.member_count}条记忆) ${scoreIndicator}`);
       sections.push(`> ${cs.summary}`);
       if (cs.key_memories.length > 0) {
@@ -279,7 +305,7 @@ function buildClusteredContextBlock(clustered: import("./client.js").ClusteredRe
             ? ` [rel:${mem.relations.map((rel) => rel.target_id).join(",")}]`
             : "";
           const importanceBar = mem.importance >= 0.7 ? "●" : mem.importance >= 0.4 ? "◐" : "○";
-          const content = truncate(mem.content, maxContentLength);
+          const content = truncate(mem.content, clusterMaxLen);
           sections.push(`- ${importanceBar}${idTag}${relTag} ${content}`);
         }
       }
@@ -287,13 +313,18 @@ function buildClusteredContextBlock(clustered: import("./client.js").ClusteredRe
   }
 
   if (clustered.standalone_memories.length > 0) {
+    const standaloneBudget = clustered.cluster_summaries.length === 0
+      ? budget
+      : Math.floor(budget * 0.3);
+    const standaloneMaxLen = Math.min(maxContentLength, Math.max(minItemChars, Math.floor(standaloneBudget / clustered.standalone_memories.length)));
+
     sections.push("\n## 📌 补充信息");
     for (const mem of clustered.standalone_memories) {
       const idTag = mem.id ? ` [id:${mem.id}]` : "";
       const relTag = mem.relations && mem.relations.length > 0
         ? ` [rel:${mem.relations.map((rel) => rel.target_id).join(",")}]`
         : "";
-      const content = truncate(mem.content, maxContentLength);
+      const content = truncate(mem.content, standaloneMaxLen);
       sections.push(`-${idTag}${relTag} ${content}`);
     }
   }
@@ -316,7 +347,6 @@ export function autoRecallHook(client: CerebroClient, containerTags: string[], t
   const phase2Multiplier = config.recall?.phase2Multiplier ?? 2;
   const llmMaxEval = config.recall?.llmMaxEval ?? 15;
   const refineStrategy = config.recall?.refineStrategy ?? "balanced";
-  const refineMediumChars = config.recall?.refineMediumChars ?? 200;
   const maxContentLength = Math.max(MIN_CONTENT_LENGTH, config.content?.maxContentLength ?? 500);
   const maxContentChars = Math.max(MIN_CONTENT_CHARS, config.content?.maxContentChars ?? 30000);
   const toastDelayMs = config.ui?.toastDelayMs ?? 7000;
@@ -333,7 +363,7 @@ export function autoRecallHook(client: CerebroClient, containerTags: string[], t
     if (policy === "none") return;
 
     try {
-      logDebug("autoRecallHook start", { sessionId: input.sessionID, agentId, policy, similarityThreshold, maxRecallResults, fetchMultiplier, topkCapMultiplier, mmrJaccardThreshold, mmrPenaltyFactor, phase2Multiplier, llmMaxEval, refineStrategy, refineMediumChars });
+      logDebug("autoRecallHook start", { sessionId: input.sessionID, agentId, policy, similarityThreshold, maxRecallResults, fetchMultiplier, topkCapMultiplier, mmrJaccardThreshold, mmrPenaltyFactor, phase2Multiplier, llmMaxEval, refineStrategy });
       const messages = sessionMessages.get(input.sessionID) ?? [];
       const userMessages = messages.filter((m) => m.role === "user");
 
@@ -399,7 +429,6 @@ export function autoRecallHook(client: CerebroClient, containerTags: string[], t
           phase2_multiplier: phase2Multiplier,
           llm_max_eval: llmMaxEval,
           refine_strategy: refineStrategy,
-          refine_medium_chars: refineMediumChars,
         },
         directory || process.env.OMEM_PROJECT_DIR,
       );
@@ -494,20 +523,14 @@ export function autoRecallHook(client: CerebroClient, containerTags: string[], t
       if (budgetRemaining < 0) {
         logDebug("autoRecallHook budget overflow", { profileChars, maxContentChars, deficit: -budgetRemaining });
       }
-      const itemCount = clustered 
-        ? (clustered.cluster_summaries.length + clustered.standalone_memories.length)
-        : newResults.length;
-      const dynamicMaxContentLength = itemCount > 0
-        ? Math.min(maxContentLength, Math.max(MIN_ITEM_CONTENT_CHARS, Math.floor(budgetRemaining / itemCount)))
-        : maxContentLength;
       logDebug("autoRecallHook budget", { 
-        maxContentChars, profileChars, budgetRemaining, itemCount, 
-        configuredMax: maxContentLength, dynamicMax: dynamicMaxContentLength 
+        maxContentChars, profileChars, budgetRemaining,
+        configuredMax: maxContentLength,
       });
 
       const block = clustered 
-        ? buildClusteredContextBlock(clustered, dynamicMaxContentLength)
-        : buildContextBlock(newResults, dynamicMaxContentLength);
+        ? buildClusteredContextBlock(clustered, budgetRemaining, maxContentLength, MIN_ITEM_CONTENT_CHARS)
+        : buildContextBlock(newResults, budgetRemaining, maxContentLength, MIN_ITEM_CONTENT_CHARS);
       if (block) {
         appendToSystem(output.system, block);
         appendToSystem(output.system, FETCH_POLICY);