fix(fs-search): keep broad glob samples representative

Remove the duplicate model-facing list tool from this branch; directory orientation remains available through bash ls. Keep the glob sampling fix, add a real ACP composition snapshot, and narrow the decision record to the shipped bug fix.

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.i18n.yaml

@@ -1,6 +1,6 @@
 # Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
-#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-27-directory-listing-tool.md
-2026-07-27-directory-listing-tool.md: fd8e8fb22d0b75a4d457e800296fa3ce2c2b5bd5
-2026-07-27-directory-listing-tool.zh.md: 7d4e797878858ea1ca09f1eea6c6a3953c9abae6
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.md
+2026-07-27-glob-sampling.md: c583f0cd8110684a94d0cd04f5cf2ae861ce7aa3
+2026-07-27-glob-sampling.zh.md: 97025a7714237fc716dee745a858140cd783ba2d

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.md

@@ -0,0 +1,45 @@
+# Agent Note: Sample over-cap glob results across the tree
+
+Status: implemented
+
+English | [中文](2026-07-27-glob-sampling.zh.md)
+
+## Problem
+
+Asked what a workspace contained, an agent described one subfolder as if it were the whole project. The workspace held 22 top-level entries and 11,485 files. `glob {"pattern":"*"}` matched 10,030 paths, but all 100 inline paths sat under one recently unpacked subtree, so the model never saw the other 21 entries.
+
+Three individually valid behaviors composed into the false impression. A glob without `/` matches basenames at any depth, so `*` means every file in the tree rather than the shell's current-directory expansion. Ripgrep's `--sort=modified` is ascending, so an archive's restored old timestamps put that subtree first. The inline page then took the head of that order without saying that it represented only one concentrated slice.
+
+## Decision
+
+A result that fits within `globMaxResults` remains complete and byte-for-byte modification-time ordered. An over-cap result is sampled round-robin across the complete result's top-level entries: every entry receives one slot before any receives a second, exhausted groups drop out, and relative order remains stable within each group. Grouping is relative to the actual search root, including an explicit `path`.
+
+The footer states that the page is a cross-entry sample rather than the modification-time head, reports how many top-level entries it reaches when that fact adds information, and preserves the complete sorted list in the spill artifact. When more top-level entries exist than inline slots, it tells the model to narrow `path`.
+
+The prompt and schema also state that a pattern without `/` matches at any depth and that glob returns files, never directory entries. Directory orientation remains ordinary shell work in deployments that expose the model-facing bash tool: use `ls` for one directory, and glob for a named file-path pattern across the tree. `ctx.fs.listDir` remains an internal provider primitive used by skill discovery; this decision adds no model-facing `list` tool.
+
+## Alternatives considered
+
+**Keep the modification-time head and only warn about concentration.** Rejected after measuring the failure shape. A warning asks the model to distrust the only paths it received; representative data fixes the answer directly.
+
+**Sample every result.** Rejected. A complete result loses nothing to truncation, so modification-time order remains useful for age-oriented questions. Sampling begins only when the head stops describing the whole.
+
+**Switch to newest-first order.** Rejected. It merely changes which concentrated subtree can dominate and removes the existing oldest-first contract without making a capped page representative.
+
+**Sample only past a skew threshold.** Rejected. No current evidence supports a deployment-wide threshold, and the model could not know which ordering contract applied. The existing cap is the explainable transition.
+
+**Balance recursively below the top level.** Deferred. First-segment balance fixes the observed failure; deeper balancing needs an unsupported depth-versus-breadth policy.
+
+**Add a model-facing `list` tool.** Rejected after implementation review. The default coding composition already exposes general bash and the model understands `ls`; a duplicate tool would add permanent schema/prompt tokens plus ordering, pagination, symlink, escaping, UI, and snapshot contracts without a distinct security or policy benefit. Thin deployments without a model-facing bash tool do not gain directory orientation from this change.
+
+**Reject `*` or silently anchor separator-free patterns.** Rejected. The same basename-at-any-depth behavior makes `*.ts` useful across a tree. Documenting the rule preserves working ripgrep semantics.
+
+## Consequences
+
+An over-cap glob page no longer answers age-order questions from its inline paths; its footer says so, and the spill artifact retains the complete sorted view. Sampling balances only the first segment beneath the search root, so a deeper hot subtree can still dominate within one top-level entry.
+
+The tool surface does not grow. The fix changes glob's prompt, schema description, canonical output (`root` records the sampling basis), and over-cap Native rendering while leaving fitting results unchanged.
+
+## Testing
+
+Package tests pin concentrated and flat results, explicit roots, more groups than the JavaScript argument limit, exhausted groups, fewer slots than groups, and paths outside the workdir. The `fs-glob-sampling` ACP scenario boots the real Loader/app/sandbox-bash composition and executes the real search plugin against a deterministic `rg` process fixture; its result spans four top-level entries instead of returning one subtree's head.

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.zh.md

@@ -0,0 +1,45 @@
+# Agent Note: 跨目录树采样超出上限的 glob 结果
+
+Status: implemented
+
+[English](2026-07-27-glob-sampling.md) | 中文
+
+## 问题
+
+用户询问工作区包含什么内容时，一个 agent（智能体）把某个子文件夹描述成了整个项目。该工作区有 22 个顶层条目和 11,485 个文件。`glob {"pattern":"*"}` 匹配到 10,030 条路径，但内联显示的 100 条路径全部位于一棵近期解压的子树中，因此模型完全没有看到其余 21 个条目。
+
+三个单独看都合理的行为叠加后造成了错误印象。不含 `/` 的 glob 会匹配任意深度的文件名，因此 `*` 表示目录树中的每个文件，而不是 shell 对当前目录执行的展开。Ripgrep 的 `--sort=modified` 按升序排列，因此归档包还原出的旧时间戳会让该子树排在最前。随后，内联页面直接截取这一顺序的前部，却没有说明它只代表集中于一处的切片。
+
+## 决策
+
+未超过 `globMaxResults` 的结果仍保持完整，且按修改时间排序的内容逐字节不变。超过上限时，系统会在完整结果的顶层条目之间按轮转方式采样：每个条目都先获得一个位置，之后才有条目获得第二个位置；已经用尽的分组会退出轮转；各组内部的相对顺序保持稳定。分组始终以实际搜索根为基准，显式指定 `path` 时也如此。
+
+footer 会说明当前页面是跨条目的样本，而不是按修改时间排序的前部；当顶层条目覆盖数能提供额外信息时，还会报告该数量；完整排序列表仍保存在 spill 产物中。若顶层条目数量超过内联位置数，footer 会要求模型缩小 `path`。
+
+提示词与 schema 还会说明：不含 `/` 的模式会匹配任意深度，glob 只返回文件，绝不返回目录条目。在向模型暴露 bash 工具的部署中，目录定位仍由普通 shell 操作完成：查看一个目录使用 `ls`，跨目录树按指定文件路径模式查找则使用 glob。skill（技能）发现流程仍将 `ctx.fs.listDir` 作为内部提供方原语使用；本决策不会新增面向模型的 `list` 工具。
+
+## 考虑过的替代方案
+
+**保留按修改时间排序的前部，只警告结果过于集中。** 测量实际故障形态后否决。警告只会要求模型怀疑自己拿到的唯一一批路径；具有代表性的数据能直接修正答案。
+
+**对所有结果采样。** 否决。完整结果没有因截断损失任何信息，因此按修改时间排序仍有助于回答关注新旧时间的问题。只有当截取前部已经无法描述整体时，才开始采样。
+
+**改为最新优先排序。** 否决。这只会改变哪一棵结果集中的子树可能占据主导；既取消了现有的最旧优先契约，也没有让受限页面更具代表性。
+
+**仅在偏斜超过阈值时采样。** 否决。目前没有证据支持适用于所有部署的统一阈值，模型也无法判断当前采用的是哪一种排序契约。现有上限是可以清楚解释的切换点。
+
+**在顶层以下递归平衡。** 暂缓。按第一路径段做平衡已经修复观测到的故障；更深层的平衡需要一套尚无依据的深度与广度取舍策略。
+
+**新增面向模型的 `list` 工具。** 实现评审后否决。默认编程组合已经提供通用 bash，模型也理解 `ls`；重复工具会永久增加 schema 与提示词所占的 token，并引入排序、分页、符号链接、转义、UI 与快照契约，却没有独立的安全或策略收益。不向模型提供 bash 工具的精简部署也不会因本次改动获得目录定位能力。
+
+**拒绝 `*`，或在不含分隔符的模式前静默加上根目录锚点。** 否决。同样的「在任意深度匹配文件名」行为使 `*.ts` 可以有效地跨目录树搜索。记录这条规则能够保留正常工作的 Ripgrep 语义。
+
+## 影响
+
+超过上限的 glob 页面无法再根据内联路径回答按时间判断新旧的问题；footer 会明确说明这一点，spill 产物仍保留完整的排序视图。采样只平衡搜索根下的第一路径段，因此某个顶层条目内部较深处、结果密集的子树仍可能占据主导。
+
+工具接口不会扩大。此修复会更改 glob 的提示词、schema 描述、规范输出（`root` 记录采样基准）以及超过上限时的 Native 渲染，未超过上限的结果保持不变。
+
+## 测试
+
+包测试锁定了结果集中与扁平两种情况、显式根目录、分组数超过 JavaScript 参数个数上限、分组耗尽、位置数少于分组数，以及工作目录以外的路径。`fs-glob-sampling` ACP（Agent Client Protocol）场景会启动真实的 Loader/app/sandbox-bash 组合，并让真实搜索插件对接确定性的 `rg` 进程 fixture（测试前置数据）；其结果覆盖 4 个顶层条目，而不是只返回某棵子树的前部。

.agents/notes/implemented/feature/2026-07-27-directory-listing-tool.md

@@ -1,109 +0,0 @@
-# Agent Note: Sample over-cap glob results across the tree, and give the model a directory-listing tool
-
-Status: implemented
-
-English | [中文](2026-07-27-directory-listing-tool.zh.md)
-
-## Problem
-
-Asked what a workspace contained, an agent described one subfolder as if it were the whole project.
-
-The session log shows exactly how. The workspace held 22 top-level entries and 11,485 files. The model called `glob {"pattern": "*"}`, which matched 10,030 paths; the tool showed the first 100, and all 100 sat under a single recently-unpacked subdirectory holding 355 of those files. The model never saw the other 21 top-level entries and answered from the one it did see. The session cwd was correct throughout — nothing was misconfigured, and every number the tool printed was true.
-
-Three properties of `glob` compose into that page:
-
-- **A pattern with no `/` matches at any depth.** The pattern goes to ripgrep as `--glob=<pattern>`, where a glob without a separator matches the basename anywhere in the tree. `*` therefore means "every file in the workspace", not "the top level" — the opposite of what it means in a shell. The tool said nothing about this, and every example in its schema was `**/…`, so nothing suggested the plain form was recursive.
-- **`--sort=modified` orders oldest first.** Unpacking an archive restores the timestamps stored inside it, which predate everything the user wrote, so a freshly unpacked subtree lands at the very front of any broad match. (Ripgrep sorts ascending; `--sortr` is the descending form and is not used here.)
-- **The inline page was the head of that order.** `globMaxResults` (100) paths were kept from the front. Under the first two properties, one subtree takes every slot.
-
-Each property is defensible alone. Together they make the most ordinary request an agent receives — "what is in this directory" — reliably produce a confident wrong answer, because nothing in the result distinguishes "the 100 oldest files in this workspace" from "this workspace".
-
-### What ordering can and cannot fix
-
-A directory's name reaches the model only as the prefix of one of its files' paths, since `rg --files` emits files and never directory entries. That is enough for ordering to matter a great deal, and not enough for `glob` to answer the question.
-
-Measured on a reproduction of the failure's shape — 24 top-level entries, 716 files, one old-timestamped subtree:
-
-| First 100 paths chosen by | Distinct top-level names visible |
-| --- | --- |
-| modification time, oldest first (the shipped behavior) | 7 |
-| round-robin across top-level entries | 21 |
-
-So a differently chosen page does surface most of the missing names, and the original diagnosis that ordering could not have helped was wrong. What no ordering fixes: an entry with no files beneath it never appears at all (the reproduction's empty directory is absent from the complete 716-path output), and nothing in the output says which names are directories or how many entries a directory holds. `glob` can therefore convey a tree's rough shape; it cannot state a directory's contents.
-
-## Decision
-
-Two changes, in the two packages that own the two halves of the failure.
-
-### The inline page of an over-cap `glob` result is sampled, not taken from the head
-
-`@deepseek-ai/dsh-tool-fs-search`. A result within `globMaxResults` is unchanged: shown whole, in modification-time order. Only when the result is larger does the page change — and there, taking the head is what fails.
-
-`sampleAcrossTopLevel` removes the displayed search-root prefix, groups the complete result by the next path segment, and fills the page round-robin: every entry immediately beneath the actual relative or absolute root gets a slot before any entry gets a second, and an entry that runs out of paths drops out so its remaining slots go to the rest. Sort order survives where it still carries meaning — groups are visited in the order ripgrep first emits them, and each group's own paths keep their relative order. The page is emitted grouped by entry rather than interleaved, so the breadth is legible at a glance.
-
-The footer states the basis, because a page that silently stopped being "the first N in modification-time order" would be a second, quieter version of the same lie:
-
-```
-(Showing 100 of 10030 paths, sampled across 22 of the 22 top-level entries this pattern matched
-instead of taken in modification-time order. Full sorted result stored at: …)
-```
-
-When the page cannot reach every top-level entry — more entries than slots — the footer says so with the shown/total spread and tells the model to narrow `path`. A flat result, where every path is its own top-level entry, keeps the plain `(Showing k of n paths. …)` footer: there the round-robin *is* the sorted head, and naming a spread would only restate the counts. The spill artifact always holds the complete list in modification-time order, so the sorted view is never lost.
-
-The guidance and schema stop misleading in the same change: "not shell find or ls" becomes "not shell find"; both now state that a pattern without `/` matches basenames at any depth, that results are files and never directories, and that a fitting result is modification-time ordered while a larger one is sampled across top-level entries. They do not recommend sibling-package tools that may be absent from the current composition.
-
-### `list`, in `@deepseek-ai/dsh-tool-fs`
-
-A fourth model-facing filesystem tool over the existing `ctx.fs.listDir` primitive, which until now shipped with skill discovery as its only consumer; [the seam Agent Note](../../archived/architecture/2026-07-03-filesystem-directory-listing-seam.md) deferred the model-facing consumer to a separate decision, which this is.
-
-It takes optional `path` and 1-based `offset` arguments, defaulting to the calling agent's session workspace and entry 1, and returns one bounded page as `{ path, offset, entries: [{ name, type }], totalEntries, counts }`, where `type` is `file`, `directory`, or `other`. It calls `listDir` and nothing else: the seam already reports absence as `FS_NOT_FOUND` and a non-directory target as `FS_NOT_DIRECTORY`, so a preceding `stat` would add a round-trip and a second source of truth.
-
-Three presentation rules carry the decision:
-
-- **Directories sort first, then files, then non-regular children, each alphabetically** before paging, so every offset traverses one stable order and the first page keeps navigable structure.
-- **The canonical value and Native result carry one recoverable page** of at most `listMaxEntries` (default 200, configurable). The footer states the complete size and composition and gives `offset=<next>` until the final page, so omitted sibling names remain reachable.
-- **Filesystem text cannot forge presentation structure.** A name is emitted verbatim unless it would make the listing lie — a control character splitting one entry across lines, `</` closing the envelope, a trailing `@` colliding with the non-regular marker — and is then a JSON string with `</` neutralized. Quoting only when needed keeps the tool an agent reaches for first legible in every transcript, while leaving the format unambiguous.
-
-`list` emits no `fs/observed`. Seeing a filename is not reading a file, and a listing must never satisfy the read-before-write gate that `@deepseek-ai/dsh-fs-policy` enforces. It declares `isConcurrencySafe`, because it mutates nothing at all.
-
-### Why both
-
-They answer different questions and neither substitutes for the other. `list` answers "what is here" exactly — entry names, their types, the complete count — which `glob` cannot do at any ordering. Sampling fixes the page a broad `glob` returns for every *other* question, which stays wrong even once a better tool exists, because the model has no reason to abandon a page that looks representative.
-
-## Alternatives considered
-
-**Leave `glob` ordering alone and only warn in the footer.** This is what the first implementation did: keep the oldest-first head and add a clause saying the head covered 1 of 22 top-level entries. Rejected once measured. A warning asks the model to distrust the only data it has and go elsewhere; a better page just is not wrong. The warning also does nothing for a model that stops reading at the paths, which is the failure being fixed.
-
-**Sample always, replacing modification-time order outright.** Rejected. Over a complete result the order answers age questions — what is stale, what was touched last — a genuinely useful and separate purpose, and a complete result is the case where the order costs nothing and means everything. Sampling only past the cap is the point where the order has already stopped describing the result: the head of a 10,030-path list is not "the oldest files worth knowing about", it is an arbitrary 1% of them.
-
-**Sample by a skew threshold — head unless the head is badly concentrated.** Rejected. A threshold is a deployment-varying tunable with no evidence behind any value, and it makes the result's ordering contract conditional on data the model cannot see. "Over the cap" is a boundary the model already knows about from the footer.
-
-**Balance recursively, not just at the top level.** Deferred, and recorded as a Known Limitation. Top-level balance fixes the observed failure and is explainable in one sentence of tool description; per-level balancing needs a policy for how depth trades against breadth, which no current evidence settles.
-
-**Reject `*`, or silently rewrite it to a top-level-anchored pattern.** Rejected. The same basename-at-any-depth rule that makes `*` recursive is what makes `*.ts` mean "every TypeScript file", the overwhelmingly common and correct use; anchoring one and not the other is an arbitrary special case, and rejecting a pattern ripgrep accepts turns a working call into an error. Documenting the rule costs nothing and generalizes.
-
-**Add `list` without touching `glob`.** Rejected, for the reason stated under *Why both* above. The misleading page is reachable from any broad pattern, and a model that believes its sample is representative has no reason to reach for another tool.
-
-**Fix `glob` without adding `list`.** Rejected for the same reason in reverse. A sampled page shows most top-level *names*, but not which are directories, not the empty ones, and not the entry count; "what is in this directory" deserves a tool that answers it rather than a sample the model must infer from.
-
-**Report per-directory child counts in `list`.** Rejected. Counting each entry's children means one `listDir` per child — an N+1 fan-out across a seam that may be remote or sandboxed, paid on every listing, to sharpen a decision the model can settle by listing the one subdirectory it cares about.
-
-**Make `list` recursive with a depth argument.** Rejected for now. One level composes: the model lists what it needs to descend into. Recursion reintroduces the size and truncation problems this note exists to fix, and the provider primitive is deliberately one-level.
-
-**Spill a capped listing through `ctx.spillStore`, as `glob` does.** Rejected for v1. `glob` needs spill because its schema has no offset; `list` has a cheap successor call through the exact next offset in the footer, while every page repeats the complete size and composition.
-
-## Consequences
-
-An over-cap `glob` result no longer returns the oldest paths at the head of modification-time order. That is a real contract change on a hot path, and it is why the footer, the prompt section, and the schema all state the sampled basis rather than leaving the model to infer it. A result within the cap is byte-identical to before, so age-ordered reading keeps working wherever it was working.
-
-Balancing is by first path segment only, so a result concentrated deeper — one enormous directory inside an otherwise even tree — is still shown unevenly below the top level. Recorded in the package's Known Limitations.
-
-The shipped tool surface grows by one tool in every deployment that loads `@deepseek-ai/dsh-tool-fs`, which is all of them: a fixed schema and prompt cost on every request, and an invalidated pinned request header in the ACP snapshot scenario that pins full system-prompt and tool-schema content. The gain is that the harness's most common question has a correct answer; the previous state was not a missing convenience but a capability hole that produced confidently wrong answers.
-
-`ctx.fs.listDir` gains its first model-facing consumer, which makes its contract load-bearing for a product surface: a future remote or sandboxed backend must implement direct-child listing well enough for a model to navigate by, not merely well enough for skill discovery. The `other` type stays collapsed at the seam, so `list` cannot distinguish a symlink from a socket and marks both `@`.
-
-## Testing
-
-Package tests pin the model-visible text of both surfaces. For `glob`: sampling over a concentrated result, an explicit relative root, more top-level groups than JavaScript's argument limit, an exhausted group handing its slots on, a page smaller than the top level, absolute paths outside the workdir, and a flat result that must reproduce the sorted head. For `list`: ordering, complete composition, offset continuation and rejection, empty directories, and filesystem names containing newlines, tag text, or marker suffixes. A registry-level test asserts that a listing emits no `fs/observed` and that a following `edit` still fails `FS_NOT_OBSERVED`, so the tool cannot become an accidental read-before-write bypass. Prompt-section registration, schema registration, and HMR disposal cover the fourth tool alongside the existing three.
-
-The assembled transcript is the `fs-list` ACP scenario: a workspace whose answer is its subdirectories, where the model calls `list` with no arguments and the pinned tool result carries the directory-first envelope and its composition footer — an answer `glob` could not have produced at all.

.agents/notes/implemented/feature/2026-07-27-directory-listing-tool.zh.md

@@ -1,109 +0,0 @@
-# Agent Note: Sample over-cap glob results across the tree, and give the model a directory-listing tool
-
-Status: implemented
-
-[English](2026-07-27-directory-listing-tool.md) | 中文
-
-## Problem
-
-被问到工作区里有什么时，agent（智能体）把其中一个子目录当作整个项目描述了一遍。
-
-会话日志把过程说得很清楚。该工作区有 22 个顶层条目、11485 个文件。模型调用了 `glob {"pattern": "*"}`，匹配到 10030 条路径；工具展示了前 100 条，而这 100 条全部位于同一个新近解包的子目录下，该子目录只占其中 355 个文件。模型从未看到另外 21 个顶层条目，只能依据它看到的那一个作答。整个过程中会话 cwd 都是正确的——没有任何配置错误，工具打印的每个数字也都是真的。
-
-`glob` 的三项性质合成了那个页面：
-
-- **不含 `/` 的 pattern 匹配任意深度。** pattern 直接作为 `--glob=<pattern>` 交给 ripgrep，而不含分隔符的 glob 匹配树中任何位置的基名。因此 `*` 的含义是「工作区里的每个文件」，而不是「顶层」——与它在 shell 里的含义正好相反。工具对此只字未提，schema 里的示例又全是 `**/…`，没有任何线索表明朴素写法是递归的。
-- **`--sort=modified` 按从旧到新排序。** 解包压缩档会还原档案内部保存的时间戳，它们早于用户自己写下的一切，因此新近解包的子树会落在任何宽泛匹配的最前面。（ripgrep 按升序排序；降序是 `--sortr`，此处未使用。）
-- **内联页面取的是该顺序的头部。** 从最前面保留 `globMaxResults`（100）条路径。在前两项性质之下，一个子树吃掉全部位置。
-
-单看每一项都站得住。合在一起，它们让 agent 收到的最普通的请求——「这个目录里有什么」——稳定地产出一个笃定的错误答案，因为结果里没有任何信息能区分「本工作区按从旧到新顺序排在最前的 100 个文件」与「本工作区」。
-
-### 排序能修什么，不能修什么
-
-由于 `rg --files` 输出文件、从不输出目录条目，一个目录的名字只能作为其下某个文件路径的前缀抵达模型。这既足以让排序变得非常重要，也不足以让 `glob` 回答那个问题。
-
-在一份复现该失败形态的目录上实测——24 个顶层条目、716 个文件、一个时间戳较旧的子树：
-
-| 前 100 条路径的挑选方式 | 可见的顶层名个数 |
-| --- | --- |
-| 按修改时间、从旧到新（已交付的行为） | 7 |
-| 跨顶层条目轮转 | 21 |
-
-也就是说，换一种页面挑选方式确实能呈现出大部分缺失的名字，最初那句「排序帮不上忙」的诊断是错的。排序修不了的是：没有任何文件的条目根本不会出现（复现目录里的空目录在完整的 716 条输出中一次都没出现），而且输出里没有任何信息说明哪些名字是目录、某个目录有多少条目。因此 `glob` 能传达一棵树的大致形状，却说不出一个目录的内容。
-
-## Decision
-
-两项改动，分别落在承担这次失败两半责任的两个包中。
-
-### 超过上限的 `glob` 结果，内联页面改为取样而非取头部
-
-`@deepseek-ai/dsh-tool-fs-search`。未超过 `globMaxResults` 的结果完全不变：整体展示，按修改时间排序。只有结果更大时页面才改变——而恰恰在这种情况下，取头部是会失败的做法。
-
-`sampleAcrossTopLevel` 移除所显示的搜索根前缀，再按下一个路径段对完整结果分组，并以轮转方式填充页面：实际相对或绝对搜索根正下方的每个条目都先拿到一个位置，任何条目才可能拿到第二个；某个条目的路径用尽后退出，其剩余份额交给其余条目。排序在仍有意义的地方被保留下来——分组按 ripgrep 首次输出它们的顺序访问，而每个分组内部的路径保持原有相对顺序。页面按条目分组输出而非交错输出，使覆盖广度一眼可辨。
-
-footer 会说明取用依据，因为一个悄悄不再是「按修改时间排序的前 N 条」的页面，只会成为同一个谎言更安静的版本：
-
-```
-(Showing 100 of 10030 paths, sampled across 22 of the 22 top-level entries this pattern matched
-instead of taken in modification-time order. Full sorted result stored at: …)
-```
-
-当页面无法触达全部顶层条目时——条目数多于位置数——footer 会给出已触达/总计的分布，并要求模型缩小 `path`。扁平结果（每个路径各自构成一个顶层条目）保留朴素的 `(Showing k of n paths. …)` footer：此时轮转结果**就是**排序后的头部，说明分布只会重复计数。spill 产物始终保存按修改时间排序的完整列表，排序视图不会丢失。
-
-同一次改动里，指导与 schema 也不再误导：「not shell find or ls」改为「not shell find」；两处现在都写明：不含 `/` 的 pattern 匹配任意深度的基名、结果只有文件从不包含目录、未超上限的结果按修改时间排序，而更大的结果跨顶层条目取样。它们不会推荐当前组合中可能不存在的兄弟包工具。
-
-### `list`，位于 `@deepseek-ai/dsh-tool-fs`
-
-在既有的 `ctx.fs.listDir` 原语之上新增第四个面向模型的文件系统工具；该原语此前交付时唯一的消费方是 skill（技能）发现，[seam Agent Note](../../archived/architecture/2026-07-03-filesystem-directory-listing-seam.md) 把面向模型的消费方推迟为一项单独的决策，也就是本文。
-
-它接受可选的 `path` 和从 1 开始的 `offset` 参数，默认取调用 agent 的会话工作区和第 1 个条目，并返回一个有界页面，形如 `{ path, offset, entries: [{ name, type }], totalEntries, counts }`，其中 `type` 为 `file`、`directory` 或 `other`。它只调用 `listDir`：seam 已经把不存在报告为 `FS_NOT_FOUND`、把非目录目标报告为 `FS_NOT_DIRECTORY`，前置 `stat` 只会多一次往返并制造第二个真源。
-
-有三条展示规则承载了这个决策：
-
-- **先目录、再文件、最后非常规子项，各组内按字母序**，然后再分页，使每个 offset 都遍历同一稳定顺序，且第一页保留可导航的结构。
-- **规范值和 Native 结果携带一个可继续取回的页面**，最多包含 `listMaxEntries` 个条目（默认 200，可配置）。footer 会说明完整规模与构成，并在最后一页之前给出 `offset=<next>`，因此被省略的同级名称仍可取回。
-- **文件系统文本无法伪造展示结构。** 条目名默认原样输出；只有当它会让列出结果失真时才转为 JSON 字符串并中和 `</`——控制字符会把一个条目拆成两行、`</` 会关闭包络、结尾 `@` 会与非常规标记撞车。只在必要时加引号，既让这个 agent 最先使用的工具在每份 transcript（文本记录）里保持可读，又让格式无歧义。
-
-`list` 不发出 `fs/observed`。看到文件名不等于读过文件，列出绝不能满足 `@deepseek-ai/dsh-fs-policy` 施加的编辑前读取门禁。它声明 `isConcurrencySafe`，因为它完全不做任何变更。
-
-### 为什么两者都要
-
-它们回答的是不同的问题，彼此无法替代。`list` 精确回答「这里有什么」——条目名、类型、完整计数——这是 `glob` 在任何排序下都做不到的。取样修的是宽泛 `glob` 为**其他**所有问题返回的那个页面；即便有了更好的工具，那个页面依然是错的，因为模型没有理由放弃一个看起来有代表性的页面。
-
-## Alternatives considered
-
-**不动 `glob` 排序，只在 footer 里警告。** 这正是第一版实现的做法：保留从旧到新的头部，加一句「该头部只覆盖 22 个顶层条目中的 1 个」。实测之后否决。一句警告是在要求模型不信任它手上唯一的数据并另寻他处；而一个更好的页面本身就不是错的。对于读到路径就停下的模型，警告也毫无作用——而那正是本次要修的失败。
-
-**一律取样，彻底取消按修改时间排序。** 已否决。在完整结果上，该顺序回答的是与新旧有关的问题——哪些已经陈旧、哪些最后被动过——这是一个确实有用且独立的用途；而未超上限的完整结果恰恰是该顺序毫无代价、意义最大的场景。只在超过上限后取样，正好落在该顺序已经不再描述结果的那个点上：一份 10030 条列表的头部不是「最值得知道的最旧文件」，而是其中任意的 1%。
-
-**按偏斜阈值决定是否取样——头部不太集中时仍取头部。** 已否决。阈值是随部署而变的可调参数，且其取值没有任何证据支撑；它还会让结果的顺序契约取决于模型看不到的数据。而「是否超过上限」是模型本来就能从 footer 得知的边界。
-
-**递归均衡，而不只在顶层均衡。** 已延期，并记入已知限制。顶层均衡修好了已观察到的失败，且能用一句工具描述解释清楚；逐层均衡需要一套关于深度与广度如何权衡的策略，目前没有证据能定下来。
-
-**拒绝 `*`，或悄悄将其改写为锚定顶层的 pattern。** 已否决。让 `*` 递归的那条「任意深度匹配基名」规则，同样让 `*.ts` 意为「每个 TypeScript 文件」，而后者是压倒性常见且正确的用法；只锚定其一是任意的特例，而拒绝一个 ripgrep 本可接受的 pattern 则把可用调用变成错误。把规则写进文档不花任何代价，而且可以推广。
-
-**只加 `list`，不动 `glob`。** 已否决，理由见上文《为什么两者都要》。任何宽泛 pattern 都能触达那个误导性页面，而一个相信自己样本有代表性的模型，没有理由改用别的工具。
-
-**只修 `glob`，不加 `list`。** 出于对称的理由否决。取样页面能展示大部分顶层**名字**，但说不出哪些是目录、看不到空目录、也给不出条目总数；「这个目录里有什么」值得一个直接回答它的工具，而不是一份需要模型去推断的样本。
-
-**在 `list` 中报告每个目录的子项数。** 已否决。统计每个条目的子项意味着对每个子项各调用一次 `listDir`——在一个可能是远程或沙箱的 seam 上做 N+1 扇出，每次列出都要付费，只为了让模型少做一个它本可以通过列出所关心的那一个子目录就完成的判断。
-
-**让 `list` 支持递归和深度参数。** 暂时否决。单层是可组合的：模型列出它需要进入的那一层即可。递归会重新引入本 Agent Note 要解决的规模与截断问题，而且提供方原语本身就有意只做一层。
-
-**像 `glob` 那样，把达到上限的列出结果通过 `ctx.spillStore` 落盘。** v1 已否决。`glob` 需要 spill，是因为其 schema 没有 offset；`list` 可以通过 footer 中精确的下一 offset 廉价地继续调用，而且每一页都会重复完整规模与构成。
-
-## Consequences
-
-超过上限的 `glob` 结果不再返回按修改时间从旧到新排序时位于头部的那些路径。这是热路径上一项真实的契约变更，也正因如此，footer、提示词段和 schema 都写明了取样这一依据，而不是留给模型去推断。未超上限的结果与此前逐字节相同，因此按新旧顺序阅读结果的用法在原本能用的地方继续能用。
-
-均衡只按路径首段进行，因此集中在更深层的结果——一棵总体均匀的树里某个特别庞大的目录——在顶层以下仍然分布不均。已记入该包的已知限制。
-
-随产品交付的工具接口在每个加载 `@deepseek-ai/dsh-tool-fs` 的部署中都多出一个工具，而这是全部部署：每个请求都要支付固定的 schema 与提示词成本，并且会使 ACP（Agent Client Protocol）快照场景中那份钉住完整系统提示词与工具 schema 内容的请求头失效。换来的是：harness 最常见的问题终于有了正确答案；此前的状态不是缺少一项便利，而是一个会产出笃定错误答案的能力空洞。
-
-`ctx.fs.listDir` 拥有了第一个面向模型的消费方，其契约由此成为产品接口的承重部分：未来的远程或沙箱后端实现直接子项列出时，标准不再是「够 skill 发现用」，而是「够模型据以导航」。`other` 类型在 seam 处仍是合并的，因此 `list` 无法区分符号链接与套接字，两者都标记为 `@`。
-
-## Testing
-
-包测试钉住两个接口面向模型的文本。`glob` 方面：对集中结果取样、显式相对根、顶层分组数量超过 JavaScript 参数上限、某分组用尽后交出份额、页面小于顶层条目数、工作目录之外的绝对路径，以及必须复现排序头部的扁平结果。`list` 方面：顺序、完整构成、offset 续页与拒绝、空目录，以及包含换行、标签文本或标记后缀的文件系统名称。一项注册表层级的测试断言列出不会发出 `fs/observed`，且随后的 `edit` 仍以 `FS_NOT_OBSERVED` 失败，使该工具不会变成意外的编辑前读取绕过口。提示词段注册、schema 注册与 HMR（热模块替换）dispose（资源释放）覆盖了这第四个工具与既有三个工具。
-
-组装后的 transcript（文本记录）由 `fs-list` ACP 场景承担：该工作区的答案就是它的子目录，模型不带任何参数调用 `list`，被钉住的工具结果携带目录优先的包络及其构成 footer —— 这个答案 `glob` 根本无法给出。