Agent can enter an infinite loop when handling large tool results with long lines

[KDKHD]

Issue: Agent can enter an infinite loop when handling large tool results with long lines

Summary

When a tool returns a large result composed of a small number of very long lines, the agent can enter an infinite eviction/read loop.

Root Cause

The issue arises from an interaction between:

• toolTokenLimitBeforeEvict (default: 20000)
• Eviction of large tool results to the filesystem
• The read_file tool’s line-based limit (default: 2000 lines)

Detailed Explanation

1. By default, toolTokenLimitBeforeEvict is set to 20000.

2. If a tool returns a result containing very long lines (e.g. 5 lines totaling ~100,000 characters), the result exceeds the eviction threshold
   (100,000 chars > 20,000 tokens * 4 chars/token) and is evicted to the filesystem.
   See:
   https://github.com/langchain-ai/deepagentsjs/blob/main/src/middleware/fs.ts#L481

3. The agent then attempts to read the evicted content using the read_file tool.

4. read_file applies a line limit (default: 2000 lines):
   https://github.com/langchain-ai/deepagentsjs/blob/main/src/middleware/fs.ts#L193

5. Because the file contains only a few lines (e.g. 5), all ~100,000 characters are returned in a single read_file result.

6. The read_file tool result now again exceeds the eviction threshold and is evicted to the filesystem.

7. The agent repeats the process, resulting in an infinite loop:

Trace

Reproduction trace:
https://smith.langchain.com/public/5ccfa3e5-07f0-4572-b205-22ae3d3d5689/r

Possible Solutions

• Change read_file to enforce a character limit instead of (or in addition to) a line limit
• Exempt read_file results from eviction. It doesn't make sense for the read_file tool to result in evictions, as this would result in files being duplicated.
• Alternatively, cap the returned content from read_file based on token or character count to guarantee forward progress