Feature request: terminal_search tool for locating lines in large terminal scrollback

[ystyle]

Problem

terminal_read pages retained scrollback by newest-relative offset and count. With backend log output running to tens of thousands of lines, a model cannot tell where the interesting lines are: it has to page backward from the newest output one page at a time, and totalLines alone does not locate a needle in the retained buffer. The tool surface has pagination but no way to find lines by content.

Proposed feature

A new model-facing terminal_search tool that searches retained scrollback by regular expression:

• Returns matching lines newest first, each with a newest-relative offset that is directly usable as a terminal_read offset — the model jumps to a match's surrounding context in one follow-up call instead of blind paging.
• Returns the exact matchCount, the retained totalLines, and truncated.
• offset / limit page over matches (skip the newest matches, cap the returned page), mirroring terminal_read pagination.
• Shares read's byte bound (maxReadBytes, cutting the oldest side so the newest matches survive) and the consumer's maxResultBytes cap.

Reference implementation

A complete, tested implementation is ready on a fork branch: ystyle/deepseek-harness:feat/terminal-search (commit 92a251077c), ready to open as a PR once the contribution channel opens.

Scope: seam types TerminalSearchRequest / TerminalSearchMatch / TerminalSearchResult plus TerminalBackendSession.search() and TerminalSessionService.search() in @deepseek-ai/dsh-terminal; the regex search over retained scrollback in @deepseek-ai/dsh-terminal-bash; the terminal_search tool and its rendering in @deepseek-ai/dsh-tool-terminal. Docs, generated catalogs, package READMEs (en/zh), and an Agent Note are updated in the same change. Validation: typecheck, lint, 146 affected unit tests (including the real-PTY suite), and all 28 doc-sync gates pass.

Happy to adapt the design or split the change per the team's preference.

中文摘要

terminal_read 只能按 offset/count 从最新输出往回翻页；日志达到数万行时，模型无法定位目标行，只能盲目翻页。建议新增 terminal_search 工具：按正则表达式搜索保留的 scrollback，返回最新优先的匹配行，每行携带 newest-relative offset（可直接配合 terminal_read 跳转查看上下文），并返回精确匹配数与保留行数。完整实现、测试与文档已在 fork 分支 ystyle/deepseek-harness:feat/terminal-search（commit 92a251077c），贡献渠道开放后可随时转为 PR。