/chat 工具卡把内核「Tool result too large」英文驱逐占位原样展示；驱逐阈值 1000 token 太低

## 现象（2026-09-06 devapp 人类实测）

deep-agent 内核（`deepagents` `FilesystemMiddleware`，`apps/deep-agent-service/src/deep_agent_service/harness.py` 钉 `tool_token_limit_before_evict=1000`）把超阈值的工具结果驱逐成沙箱文件 `/large_tool_results/`，ToolMessage 正文换成给模型看的英文占位：

> Tool result too large, the result of this tool call call_xxx was saved in the filesystem at this path: /large_tool_results/call_xxx … You can read the result from the filesystem by using the read_file tool …

/chat 工具卡（`apps/web/components/chat/copilotkit-v2-tool-renderers.tsx`）把这段原样展示给用户；`task` 子代理的汇报稍长就被驱逐，用户看到一堆内部搬运痕迹。人类反馈"这个结果看起来也挺奇怪的"。

## 处置（ad-hoc，按 `.harness/instructions/ad-hoc-fix-pr-sop.md`）

1. 前端：纯函数 `apps/web/lib/tool-result-eviction.ts` 识别该形状（以 `Tool result too large` 开头且含 `/large_tool_results/`），工具卡改成一句中文「结果较长，已存为文件 …，agent 会按需读取」+ 可折叠「查看原文」；vitest 单测。
2. 后端：阈值默认提高到 8000，可由 `KERNEL_DEEP_AGENT_TOOL_EVICT_TOKENS` 覆盖；pytest 通过 env 压小阈值保持机制可测。

Start with apps/web/components/chat/copilotkit-v2-tool-renderers.tsx and apps/web/lib/tool-result-eviction.ts, then run the related Vitest tests. Inspect apps/deep-agent-service/src/deep_agent_service/harness.py and the existing pytest coverage for the eviction threshold and environment override. Done means the UI presents a localized, collapsible result notice and the backend default and override behavior are covered by tests.

Written by the indexing model from the issue text.