From 78292e836f40362c2c687598f5c9b7330131f352 Mon Sep 17 00:00:00 2001
From: Lily Miller <lilymiller@ucar.edu>
Date: Mon, 6 Oct 2025 16:59:09 -0600
Subject: [PATCH] Improve tool output processing: raise pass-through threshold,
 use extraction for medium outputs

FIXES:
1. Truncation was too aggressive (2KB threshold, truncate to 2000 chars)
2. Important data was being lost in medium-sized outputs (2-10KB)

Changes:
- Raise pass-through threshold from 2KB to 5KB
- Medium outputs (5-20KB) now use hierarchical extraction instead of truncation
- _extract_key_findings already handles chunking automatically
- Better preservation of important data like service lists

Benefits:
- Full service lists will now be properly analyzed
- No more missing services due to truncation
- Macha can see the complete picture before responding
---
 agent.py | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/agent.py b/agent.py
index 5398e19..0246f25 100644
--- a/agent.py
+++ b/agent.py
@@ -683,26 +683,22 @@ Provide unified summary (max 800 chars) covering all key points."""
     def _process_tool_result_hierarchical(self, tool_name: str, result: Any) -> str:
         """
         Intelligently process tool results based on size:
-        - Small (< 2KB): Pass through directly
-        - Medium (2-10KB): Truncate with head+tail
-        - Large (> 10KB): Hierarchical extraction in separate context
+        - Small (< 5KB): Pass through directly
+        - Medium (5-20KB): Hierarchical extraction with single-pass summarization
+        - Large (> 20KB): Hierarchical extraction with chunked processing
         """
         result_str = json.dumps(result) if not isinstance(result, str) else result
         size = len(result_str)
         
         # Small outputs: pass through directly
-        if size < 2000:
+        if size < 5000:
             print(f"  [Tool result: {size} chars, passing through]")
             return result_str
         
-        # Medium outputs: truncate with head+tail
-        elif size < 10000:
-            print(f"  [Tool result: {size} chars, truncating to 2000]")
-            return self._simple_truncate(result_str, 2000)
-        
-        # Large outputs: hierarchical extraction
+        # Medium and large outputs: hierarchical extraction with chunking
         else:
             print(f"  [Tool result: {size} chars, extracting key findings...]")
+            # _extract_key_findings automatically chunks large outputs
             return self._extract_key_findings(tool_name, result_str)
     
     def _prune_messages(self, messages: List[Dict], max_context_tokens: int = 80000) -> List[Dict]: