edit/write results give the model zero visibility into what actually changed — the diff exists in the log but never reaches the model

[wulun811]

I'm evaluating dsh as a serious coding-agent runtime and noticed something that seems like a blind spot worth discussing (I checked the source before posting, so this is grounded):

Observation. After a successful edit or write, the model receives only a one-line acknowledgement — The file has been updated successfully. (see formatEditOutput in packages/fs/tool-fs/src/edit.ts). The full before/after diff is computed (computeHunkDiffs, with context lines) and attached as presentationMeta — persisted to the session log and rendered as the UI diff card — but it is never rendered into the model's content. The README states this explicitly: "The full write or replacement text remains in the assistant tool-call arguments."

Why this is a real blind spot, not a nit:

replace_all: true can replace more (or fewer) occurrences than the model expected; the model has no signal that the actual mutation diverged from intent.
The diff basis is LF-normalized (before/after from the backend), so files with CRLF or odd whitespace can produce replacement behavior that surprises the model.
In general: the model knows what it said (its own args) but never what the world became. Its only self-verification path is a follow-up read or bash git diff — an extra round trip it must remember to make.
Why it should be cheap to fix: the diff is already in the session log (presentationMeta → meta.diffs). Rendering a bounded version into model content changes no invariant — "model-visible ⟺ logged" holds either way. The cost concern is real (tokens + append-only KV-cache growth until compaction), which is presumably why the Claude Code-style acknowledgement was chosen.

Options I'd love input on:

Include a bounded compact diff (e.g. N hunks / M lines, capped) in edit/write result content — maybe only when multiple hunks or replace_all was used.
A deployment-configurable toggle (per the "no hardcoded tunables" convention).
Keep results as-is and document a "verify-read contract" instead.
Bonus observation for the plugin ecosystem: the pipeline already makes this patchable from outside — at tools/post-execute, result.meta carries the persisted diffs (computed in createSuccessResult before the waterfall), and a listener can replace the result content. So a community plugin can give the model the diff today without any core change.

Is this working as intended, or is "the model never sees what changed" a known trade-off you'd like to revisit?
我在认真评估把 dsh 用作正式编码 Agent 运行时，注意到一个疑似盲区，想公开讨论一下（发帖前已核对源码，以下结论都有出处）：

现象。 edit 或 write 成功后，模型只收到一句确认——The file has been updated successfully.（见 packages/fs/tool-fs/src/edit.ts 的 formatEditOutput）。完整的 before/after diff 其实是被算出来的（computeHunkDiffs，带上下文行），作为 presentationMeta 挂上——落进会话日志、渲染成 UI 的 diff 卡片——但从不渲染进模型的内容。README 里写得很直白："完整的新内容留在模型自己的工具调用参数里"。

为什么这是真盲区，不是吹毛求疵：

replace_all: true 可能替换了比模型预期更多（或更少）的位置；实际变更与意图偏离时，模型得不到任何信号；
diff 基底是 LF 归一化后的 before/after，遇到 CRLF 或怪异空白的文件，替换行为可能让模型意外，而它无从察觉；
一句话：模型知道自己说了什么（它自己的参数），却永远不知道世界变成了什么。它唯一的自证路径是事后补一次 read 或 bash git diff——一个它得记得去做、且多花一整轮的往返。
为什么修起来应该很便宜： diff 已经在会话日志里了（presentationMeta → meta.diffs）。把有上限的版本渲染进模型内容，不碰任何不变量——"模型可见 ⟺ 已记录"两边都成立。成本顾虑是真实的（token 消耗 + 追加式 KV 缓存到压缩前的持续计费），大概也正是当初选择 Claude Code 式一句话确认的原因。

想听大家意见的几个选项：

在 edit/write 结果内容里附带有上限的紧凑 diff（比如 N 个 hunk / M 行封顶）——或许只在多 hunk 或用了 replace_all 时附带；
做成部署可配置开关（符合项目"无硬编码可调项"的约定）；
结果保持现状，改为文档化一个"改后自验（verify-read）契约"。
给插件生态的一个附赠观察： 这条管道从外部就能打补丁——在 tools/post-execute 时，result.meta 已携带持久化的 diffs（在瀑布之前由 createSuccessResult 算出），而监听器可以替换结果内容。也就是说，社区插件今天就能让模型看到 diff，完全不用改核心。

这属于"按设计如此"，还是"模型永远看不到变更结果"这个取舍你们其实也想重新议一议？