@comfanion/usethis_search 4.3.0-dev.0 → 4.3.0-dev.2

package/hooks/message-before.ts

@@ -57,7 +57,9 @@ export function createWorkspaceInjectionHandler(state: SessionState) {
     // Don't inject or prune for sub-agents (title generation, etc.)
     if (state.isSubAgent) return
 
-    // ── Prune: replace old tool outputs with compact summaries ────────────
+    // ── Prune & Compact: optimize chat history ────────────────────────────
+    // 1. Prune: replace old tool outputs with compact summaries
+    // 2. Compact: remove old tool calls entirely (keep last N turns)
     // Files are already in workspace injection — no need for big outputs
     // in chat history. This runs even when workspace is empty
     // (handles case where workspace was cleared but old outputs remain).
@@ -65,6 +67,7 @@ export function createWorkspaceInjectionHandler(state: SessionState) {
     if (wsConfig.autoPruneSearch !== false) {
       pruneSearchToolOutputs(output.messages)
       pruneReadToolOutputs(output.messages)
+      compactOldToolCalls(output.messages)
     }
 
     let entries = workspaceCache.getAll()
@@ -427,3 +430,137 @@ function extractFilePathFromOutput(output: string): string | null {
   
   return null
 }
+
+// ── Tool Call Compaction ────────────────────────────────────────────────────
+
+/**
+ * Remove old tool calls (search/read) from chat history.
+ * 
+ * Strategy:
+ * - Keep last N turns (default: 5) — agent may reference recent calls
+ * - Only compact search/read tools (not edit/write/grep/glob)
+ * - Only compact completed calls with pruned outputs
+ * - Remove both call + output parts
+ * - Add compact marker at start showing how many calls removed
+ * 
+ * Why: Tool calls contain full args (200+ tokens). After pruning outputs,
+ * the calls themselves are redundant — chunks already in workspace.
+ * 
+ * Savings: ~220 tokens per compacted call × N calls = 2K-10K tokens
+ */
+const KEEP_LAST_N_TURNS = 5
+const COMPACT_TOOLS = ['search', 'read', 'Read']
+
+interface ToolCallPair {
+  msgIndex: number
+  callPart: MessagePart
+  outputPart?: MessagePart
+  tool: string
+  status: string
+  turnsSinceEnd: number
+}
+
+/**
+ * Compact old tool calls by removing them from chat history.
+ * Keeps last N turns intact.
+ */
+export function compactOldToolCalls(messages: Message[]): void {
+  // Find all tool call pairs
+  const toolPairs = findToolCallPairs(messages)
+  
+  if (toolPairs.length === 0) return
+  
+  // Calculate turns from end for each pair
+  const totalTurns = messages.length
+  
+  // Filter: only old, completed, search/read with pruned outputs
+  const toCompact = toolPairs.filter(pair => {
+    const turnsFromEnd = totalTurns - pair.msgIndex
+    return (
+      turnsFromEnd > KEEP_LAST_N_TURNS &&
+      pair.status === 'completed' &&
+      COMPACT_TOOLS.includes(pair.tool) &&
+      pair.outputPart &&
+      isPrunedOutput(pair.outputPart.state?.output || '')
+    )
+  })
+  
+  if (toCompact.length === 0) return
+  
+  // Remove tool parts from messages
+  const removedIds = new Set<string>()
+  
+  for (const pair of toCompact) {
+    removedIds.add(pair.callPart.id)
+    if (pair.outputPart) {
+      removedIds.add(pair.outputPart.id)
+    }
+  }
+  
+  // Filter out removed parts from messages
+  for (const msg of messages) {
+    if (!msg.parts || !Array.isArray(msg.parts)) continue
+    msg.parts = msg.parts.filter(part => !removedIds.has(part.id))
+  }
+  
+  // Add compact marker to first user message
+  const firstUserMsg = messages.find(m => m?.info?.role === 'user')
+  if (firstUserMsg && firstUserMsg.parts) {
+    const marker = {
+      type: 'text',
+      text: `<!-- ${toCompact.length} tool calls compacted (search/read results in workspace) -->`,
+      id: 'compact-marker-' + Date.now(),
+    }
+    firstUserMsg.parts.unshift(marker)
+  }
+}
+
+/**
+ * Find all tool call + output pairs in messages.
+ */
+function findToolCallPairs(messages: Message[]): ToolCallPair[] {
+  const pairs: ToolCallPair[] = []
+  
+  for (let i = 0; i < messages.length; i++) {
+    const msg = messages[i]
+    if (!msg.parts || !Array.isArray(msg.parts)) continue
+    
+    for (const part of msg.parts) {
+      if (part.type === 'tool' && part.tool) {
+        const status = part.state?.status || 'unknown'
+        
+        // Find matching output part (usually in same message or next)
+        let outputPart: MessagePart | undefined
+        
+        // Check same message first
+        for (const p of msg.parts) {
+          if (p.type === 'tool' && p.tool === part.tool && p.state?.output && p.id !== part.id) {
+            outputPart = p
+            break
+          }
+        }
+        
+        pairs.push({
+          msgIndex: i,
+          callPart: part,
+          outputPart,
+          tool: part.tool,
+          status,
+          turnsSinceEnd: 0, // Will be calculated in compactOldToolCalls
+        })
+      }
+    }
+  }
+  
+  return pairs
+}
+
+/**
+ * Check if output is pruned (compact format).
+ */
+function isPrunedOutput(output: string): boolean {
+  if (!output) return false
+  
+  // Pruned outputs start with [ or ✓
+  return output.startsWith('[') || output.startsWith('✓')
+}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@comfanion/usethis_search",
-  "version": "4.3.0-dev.0",
-  "description": "OpenCode plugin: semantic search with auto-attach, line numbers in workspace, simplified API (v4.3: auto-detect modes, read() caching, 99% token reduction, no grep needed)",
+  "version": "4.3.0-dev.2",
+  "description": "OpenCode plugin: semantic search with auto-attach, line numbers in workspace, simplified API (v4.3: auto-detect modes, read() caching, tool call compaction, 99% token reduction, no grep needed, LSP memory leak fixed)",
   "type": "module",
   "main": "./index.ts",
   "exports": {

package/vectorizer/graph-builder.ts

@@ -288,6 +288,18 @@ export class GraphBuilder {
     return result
   }
 
+  /**
+   * Cleanup: shutdown LSP analyzer to prevent memory leaks.
+   * MUST be called after indexing to close LSP server processes.
+   */
+  async cleanup(): Promise<void> {
+    try {
+      await this.lspAnalyzer.shutdown()
+    } catch {
+      // Best effort — don't throw if already closed
+    }
+  }
+
   // ---- FR-005: Semantic similarity edges ------------------------------------
 
   /**

package/vectorizer/index.ts

@@ -571,11 +571,15 @@ class CodebaseIndexer {
       try { this.chunkStore.close(); } catch { /* best effort */ }
       this.chunkStore = null;
     }
+    // Cleanup GraphBuilder (shutdown LSP to prevent memory leaks)
+    if (this.graphBuilder) {
+      try { await this.graphBuilder.cleanup(); } catch { /* best effort */ }
+      this.graphBuilder = null;
+    }
     // Close graph DB to release LevelDB lock
     if (this.graphDB) {
       try { await this.graphDB.close(); } catch { /* best effort */ }
       this.graphDB = null;
-      this.graphBuilder = null;
     }
     // Save & release usage tracker
     if (this.usageTracker) {
@@ -1623,6 +1627,15 @@ class CodebaseIndexer {
       }
     }
 
+    // Cleanup: shutdown LSP to prevent memory leaks after bulk indexing
+    if (this.graphBuilder) {
+      try {
+        await this.graphBuilder.cleanup();
+      } catch {
+        // Best effort — continue even if cleanup fails
+      }
+    }
+
     return { indexed, skipped, total, semanticEdges };
   }
 

package/vectorizer.yaml

@@ -26,6 +26,7 @@ vectorizer:
       min_chunk_size: 1000  # Merge small sections (avoid header-only chunks)
       max_chunk_size: 8000  # Large chunks for docs (SQL schemas, API specs, etc.)
       preserve_heading_hierarchy: true
+      skip_low_priority: true  # Skip SQL schemas, continuous aggregates (default: true)
     code:
       split_by_functions: true
       include_function_signature: true