[Feature]: 工具结果返回驱逐后导致llm上下文超限

[oyx888]

agent调用工具如果是返回一长串json时，驱逐大结果工具会往文件中一行写入，导致read_file读取文件时只读取到一行导致llm上下文超限，read_file工具是在白名单内，从而导致agent运行不了

[logicwu0]

我看了一下当前实现，问题应该在于分页维度不一致：ToolResultEvictionMiddleware 会把完整结果原样写入文件，而 read_file 的 offset / limit 只按行分页。对于单行超长 JSON，即使只读取一行，也会把整行重新放回模型上下文，因此驱逐并没有真正限制下一次读取量。

我不太建议只对 JSON 做 pretty-print：如果某个字段本身是超长字符串，或者结果不是 JSON，仍然会出现超长单行。更通用的方案是给 read_file 增加可选的字符分页参数，例如 char_offset 和 char_limit：

• 保留现有 offset / limit 的行分页语义；
• 先按行选择内容，再对返回文本应用字符窗口；
• 超出窗口时附加明确的 continuation 提示，例如下一次使用 char_offset=20000；
• 不修改驱逐文件的原始内容，保证完整结果仍可分段读取；
• 保留现有 Java 方法作为兼容重载，避免已有直接调用受影响；
• 截断时避免拆开 UTF-16 surrogate pair。

计划补充的测试包括：单行超长 JSON 的首次读取和续读、多行文件现有分页行为、短文件不截断、中文/emoji 边界，以及驱逐文件内容保持不变。驱逐后的占位提示也可以明确建议带 char_limit 调用 read_file。

请问维护者是否接受按这个“字符分页 + 保留原文件”的方向处理？如果方向认可，我可以再基于最新 main 提交一个聚焦的修复和回归测试。

[oss-maintainer]

Issue Evaluation

Category: enhancement | Status: Evaluated

This is the same underlying defect as #2951, reported from the eviction angle: when a large tool result is evicted to a file it is written as a single long line, and read_file (which paginates by line only and is excluded from eviction via DEFAULT_EXCLUDED_TOOLS) returns the entire line, re-inflating the model context.

We have independently confirmed the root cause in #2951 against current main: LocalFilesystem.read() applies no character/byte cap, and read_file is in ToolResultEvictionConfig.DEFAULT_EXCLUDED_TOOLS, so an oversized single-line result bypasses eviction. The direction discussed above (char-level pagination, e.g. char_offset/char_limit, keeping line semantics and the eviction file intact) is consistent with the fix being tracked under #2951.

Recommendation: consolidate with #2951 — the fix belongs in the filesystem read path (char/byte budget + continuation metadata), which resolves both the direct-read case (#2951) and the evicted-artifact case described here. We recommend maintainers keep this issue linked to #2951 rather than treating it as a separate change.

---

Automated evaluation by github-manager-bot