From 92cf31284333ae8331e2f086d9461d2ecf4ae700 Mon Sep 17 00:00:00 2001
From: Jiayuan Zhang <forrestchang7@gmail.com>
Date: Sun, 15 Feb 2026 21:37:15 +0800
Subject: [PATCH] refactor(compaction): remove pre-flight tool result pruning
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Pre-flight compaction runs in-memory only (not persisted), so tool
result pruning in this path was wasted work — results were thrown away
after the LLM call. Post-turn compaction still handles pruning and
persists the results. Only Phase 2 (emergency message drop) remains
as a safety net in pre-flight.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 packages/core/src/agent/runner.ts | 40 +++++--------------------------
 1 file changed, 6 insertions(+), 34 deletions(-)

diff --git a/packages/core/src/agent/runner.ts b/packages/core/src/agent/runner.ts
index da3112a1..5a3e51bd 100644
--- a/packages/core/src/agent/runner.ts
+++ b/packages/core/src/agent/runner.ts
@@ -27,9 +27,6 @@ import {
   compactMessagesTokenAware,
   MIN_KEEP_MESSAGES,
 } from "./context-window/index.js";
-import {
-  pruneToolResults,
-} from "./context-window/tool-result-pruning.js";
 import { mergeToolsConfig, type ToolsConfig } from "./tools/policy.js";
 import {
   loadAuthProfileStore,
@@ -866,37 +863,12 @@ export class Agent {
     const originalCount = messages.length;
     let result = messages;
 
-    // Phase 1: Prune tool results (soft trim + hard clear)
-    const pruneResult = pruneToolResults({
-      messages: result,
-      contextWindowTokens: this.contextWindowGuard.tokens,
-    });
-    if (pruneResult.changed) {
-      result = pruneResult.messages;
-      if (pruneResult.softTrimmed > 0 || pruneResult.hardCleared > 0) {
-        this.runLog.log("tool_result_pruning", {
-          soft_trimmed: pruneResult.softTrimmed,
-          hard_cleared: pruneResult.hardCleared,
-          chars_saved: pruneResult.charsSaved,
-          phase: "preflight",
-        });
-      }
-    }
-
-    // Re-estimate after pruning
-    const afterPrune = estimateTokenUsage({
-      messages: result,
-      systemPrompt: this.agent.state.systemPrompt,
-      contextWindowTokens: this.contextWindowGuard.tokens,
-      reserveTokens: this.reserveTokens,
-    });
-
-    // Phase 2: Drop oldest messages if still over threshold
-    if (afterPrune.utilizationRatio >= COMPACTION_TRIGGER_RATIO) {
-      const compacted = compactMessagesTokenAware(result, afterPrune.availableTokens);
-      if (compacted) {
-        result = compacted.kept;
-      }
+    // Drop oldest messages if over threshold (emergency safety net).
+    // Tool result pruning is skipped here — it's handled by post-turn
+    // compaction which actually persists the results.
+    const compacted = compactMessagesTokenAware(result, estimation.availableTokens);
+    if (compacted) {
+      result = compacted.kept;
     }
 
     if (result.length < originalCount) {