[Feature Request] Add summarization tool for handling long tool outputs

[Wendong-Fan]

Required prerequisites

• I have searched the Issue Tracker and Discussions that this hasn't already been reported. (+1 or comment there if it has.)
• Consider asking first in a Discussion.

Motivation

Background

This is a follow-up to issue #3376 and the reverted PR #3259. We've been exploring a new approach to handle long tool outputs that consume excessive tokens in the agent's context.

Problem

When agents use tools that return very long outputs (e.g., web page snapshots, large file contents, API responses), these outputs
accumulate in memory and significantly increase token consumption in subsequent interactions. The previous automatic caching
approach was reverted due to concerns about agent performance and accuracy.

Proposed Solution

Instead of automatically caching long outputs, we propose adding a new tool that allows the agent to proactively manage long tool
outputs. When the agent recognizes that a previous tool output is excessively long, it can call this tool to:

1. Generate a summary of the previous long tool output
2. Offload the original content with a unique ID for potential future retrieval
3. Collapse/replace the long output in memory with the summary + offload reference

If LLM doing paralle tool call, no additional api request is needed

Why This Approach

Aspect	Previous Approach (Auto-offload)	New Approach (Agent-driven Tool)
Control	Automatic based on threshold	Agent decides when to summarize
Information Retention	ID reference	Summary + offload original
Intelligence	Rule-based	Context-aware decision by agent
Flexibility	Fixed behavior	Agent can specify what to focus on in summary

Implementation Considerations

• Modify memory management to support replacing/collapsing previous tool outputs
• Add a companion retrieve_ offloaded_output tool for accessing original data when needed
• Well designed docstring to guide agents on when to use this tool

Related

• Issue: [Enhance] Revive Tool Output Offloading #3376
• Reverted PR: feat: add tool call caching for chatAgent #3259

Solution

No response

Alternatives

No response

Additional context

No response

[Wendong-Fan]

thanks @Tanuj-Taneja1 for taking this issue!

[Tanuj-Taneja1]

Hey @Wendong-Fan , @MuggleJinx
I wanted to ask some quick questions
First : The initial approach stored tool output both in memory (using _tool_output_history a list) and files, can we remove the file related logic as we can just fetch it from normal memory (_tool_output_history )
Second: How do we handle output of retrieve_cached_output, whenever the agent will call the retrieve_cached_output the initial output will get re-entered in the chatagent's memory

[bytecii]

Maybe we can put first several hundreds of tokens to the context directly and put the whole tool results to memory/filesystem and provide tools for agent to load all the output if needed