[FEATURE] Improve tool result truncation strategy for graceful context reduction

[kevmyung]

Problem Statement

The current SlidingWindowConversationManager truncation logic has several issues:

1. Truncates newest tool results first - The _find_last_message_with_tool_results method searches from newest to oldest and truncates the most recent tool result. This is problematic because:

   • The newest tool result is likely the most relevant to the current task
   • Agent may interpret the truncated result ("The tool result was too large!" with status: error) as a tool failure
   • This can cause the agent to repeatedly call the same tool or make incorrect decisions based on misleading error status

2. No size check before truncation - Tool results are truncated regardless of their actual size. Small tool results may be unnecessarily truncated.

3. Complete content removal - The entire tool result content is replaced, losing all context about what the result contained.

Proposed Solution

Implement a more graceful truncation strategy:

1. Process older messages first - Truncate oldest tool results before newest ones, as older results have likely already been processed and acted upon by the agent.

2. Size-based threshold - Only truncate content blocks that exceed a configurable character threshold.

3. Partial truncation with context preservation - Instead of removing entire content, preserve the beginning and end while truncating the middle:

   [first 200 chars of original content...]
   
   ... [truncated: 15000 chars removed] ...
   
   [...last 200 chars of original content]
   

4. Image placeholder for older messages - Replace images in older messages with descriptive placeholders:

   [image: filename.png, 1024x768]
   

Use Case

1. Prevent agent confusion - Agent won't misinterpret truncated recent results as failures
2. Maintain task continuity - Recent tool results remain intact for ongoing work
3. Preserve partial context - Truncated content still provides hints about what information was available
4. Reduce unnecessary retries - Agent won't repeatedly call tools thinking previous calls failed

Alternatives Solutions

Separate into dedicated ConversationManager:

• Extract truncation/compaction logic into a new CompactingConversationManager
• Keep SlidingWindowConversationManager focused solely on message count management (sliding window)
• Users can combine both managers as needed, following Single Responsibility Principle

Additional Context

I've implemented and tested this approach in a separate project with positive results (67% token reduction, improved retention).

Details: Link