feat(cli): even out the shipped tool rosters across both surfaces

The two shipped surfaces offered different tools for no recorded reason:
session checkpoints, tool-result pruning, the goal tools, and Ralph were in
`tui.cordis.yml`; `tool-todo` and web search were in `web.cordis.yml`. Neither
offered session search, a string-replacement editor, or a repeat-tool guard,
though none of the three is surface-specific.

Move the rows that are not surface-specific into `base.cordis.yml` and add
those three. Web search moves there too — the TUI decision the change that
made it a Web default deferred. Both surfaces now assemble the same 27 tools.

This adds only. No row is removed from either surface and no existing row's
configuration is edited: executors, sandbox composition, access defaults,
`tools.mode`, and the workflow tool are exactly what they were. Two rows stay
surface-specific: `tmux-context` (no terminal multiplexer in a browser) and
`session-reference` (its index has one writer owner).

Ship `dsh-mcp-client` as a dependency without a row: the plugin mounts one
server per instance and `command` is required, so a default would name a
third-party server and spawn it outside `ctx.bash` on every launch. The CLI
README carries the YAML for mounting one from a personal config.

.agents/notes/implemented/feature/2026-07-31-even-out-shipped-tool-rosters.i18n.yaml

@@ -0,0 +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/2026-07-31-even-out-shipped-tool-rosters.md
+2026-07-31-even-out-shipped-tool-rosters.md: 28b44e68228043bcccd2d5f8bef96a86ffcae9d4
+2026-07-31-even-out-shipped-tool-rosters.zh.md: 8f26f7e0f86837a235bf96fa25d74f053c48e54b

.agents/notes/implemented/feature/2026-07-31-even-out-shipped-tool-rosters.md

@@ -0,0 +1,69 @@
+# Agent Note: Even out the shipped tool rosters
+
+Status: implemented
+
+English | [中文](2026-07-31-even-out-shipped-tool-rosters.zh.md)
+
+## Problem
+
+The two shipped `dsh` surfaces offered different tools for no recorded reason. Session checkpoints, tool-result pruning, the goal tools, and Ralph were in `tui.cordis.yml`; `tool-todo` and, later, web search were in `web.cordis.yml`. Neither surface offered session search, a string-replacement editor, or a repeat-tool guard, though all three exist as packages and none is surface-specific.
+
+The result was a user-visible difference nobody had decided: the same model, asked the same thing, could set a goal on the terminal but not in the browser, and could search the web in the browser but not on the terminal.
+
+## Decision
+
+The rows that are not surface-specific move into [`base.cordis.yml`](../../../../apps/cli/config/base.cordis.yml), and three more join them: `tool-session-query`, `tool-str-replace-editor`, and `repeat-tool-guard`. Web search moves there too, which is the TUI decision that the change making it a Web default explicitly deferred. Both surfaces now assemble the same twenty-seven tools.
+
+Two rows stay surface-specific. `tmux-context` is TUI-only because a browser surface has no terminal multiplexer to describe. `session-reference` is TUI-only because it drives the shared session-query index from the launcher's process-local path, and the browser sidebar reconciles that index on its own first search.
+
+**This change adds only.** No row is removed from either surface and no existing row's configuration is edited: the executors, the sandbox composition, the access defaults, `tools.mode`, and the workflow tool are exactly what they were. A reader comparing the two catalogs before and after should find additions and nothing else.
+
+### What stays unmounted, and why
+
+Three capabilities stay out on the evidence their own packages record, and are listed here so "we forgot" and "we decided against" stay distinguishable.
+
+**`dsh-tool-cordis`** lets the model write JavaScript and mount it as a temporary plugin. Its README states the limit: "The sandbox is containment for honest code, not a security boundary — host-realm helpers on the sandbox global are reachable, so mount code can reach Node" ([Known limitations](../../../../packages/cordis/tool-cordis/README.md)). The `node:vm` realm lives inside the harness process while `dsh-sandbox-local` confines only the argv it spawns, so on the Web surface both the sandbox and the approval seam are bypassed rather than enforced.
+
+**`dsh-web-fetch-local`** stays unmounted and `dsh-tool-web` keeps `fetch: false`. SSRF protection is deferred in the implementation ([`policy.ts`](../../../../packages/web/web-fetch-local/src/policy.ts) validates protocol, credentials, and length only) and the package says so: "this provider is an SSRF primitive and **must not be enabled** in a deployment that can reach sensitive internal network targets" ([README](../../../../packages/web/web-fetch-local/README.md)). The model chooses the target, which includes the harness's own gateway on loopback, private ranges, and cloud metadata endpoints.
+
+Withholding it narrows the surface without removing the reach: `bash` is mounted, so `curl` gets the same page, as a live run confirmed. What the absence buys is the removal of an argument-shaped request primitive that needs no shell — and with it the accidental path where a summarization request quietly reaches loopback. A deployment that must contain outbound traffic needs a network-level control.
+
+**The LSP trio** stays out for an operational reason rather than a security one: `command` resolves from `PATH` at plugin load, so a missing language server fails the whole boot rather than one tool. It becomes mountable once absence degrades to a skipped registration.
+
+### MCP is a dependency, not a row
+
+`@deepseek-ai/dsh-mcp-client` becomes a runtime dependency of the CLI without a row in any shipped config. The plugin mounts exactly one server per instance and `command` is required, so a default would have to name a third-party server and spawn it as a child process on every launch — outside `ctx.bash`, and therefore outside the sandbox policy the Web surface composes.
+
+The layer that would make MCP a default is the one this repository does not have yet: a bridge that reads a user's server list and mounts one client per entry, the same shape [`dsh-hooks-claude`](../../../../packages/hooks/hooks-claude/README.md) already has for a Claude Code `hooks.json`. Shipping the dependency means an installed `dsh` can mount servers from `$DSH_HOME/config.yaml` today; the CLI README carries the YAML.
+
+## Testing
+
+[`apps/cli/tests/shipped-composition.e2e.ts`](../../../../apps/cli/tests/shipped-composition.e2e.ts) boots the shipped tree through the real Loader in a pseudo-terminal and reads the tool names out of the `request/header` the session log persisted, so the assertion is the catalog the model was actually sent. Its `--config` overlay, [`composition-keyless-tail.cordis.yml`](../../../../apps/cli/tests/fixtures/composition-keyless-tail.cordis.yml), is test isolation only: a network-free adapter and workspace-local session artifacts.
+
+That tail also inserts [`composition-settled.ts`](../../../../apps/cli/tests/fixtures/composition-settled.ts), which announces settled Loader activation on the terminal stream. The TUI renders as soon as its own fiber starts, so a prompt typed at the banner can reach the loop while tool rows and persistence are still activating and assemble a partial catalog; gating the smoke's first prompt on that marker is what makes the assertion deterministic.
+
+The same smoke pins the TUI's unchanged execution posture from the same artifact: `tool-bash` emits its `sandbox_permissions` escalation pair only when the mounted executor has wider modes to escalate to, so asserting its **absence** fails if a later change quietly sandboxes this surface.
+
+[`apps/web/tests/shipped-composition.e2e.ts`](../../../../apps/web/tests/shipped-composition.e2e.ts) covers the Web surface in the built lane, asserting its catalog, that its access default is untouched, and that `workspace-write`'s writable roots include the temp directories — a trap that makes sandbox tests lie when the workspace sits under `/tmp` ([`roots.ts`](../../../../packages/sandbox/sandbox/src/roots.ts)).
+
+`glob` and `grep` are asserted as an all-or-nothing pair rather than fixed members: `dsh-tool-fs-search` probes `command -v rg` at load and registers neither tool without ripgrep, which is a host dependency.
+
+Beyond the committed tests, both surfaces were driven against a real key from the built `apps/cli/lib/bin.js` under plain Node. Every mounted tool executed successfully, including `ralph` and `web_search`; the model never reached `cordis_*` or `mcp_*`, fell back to `grep` when asked for LSP navigation, and used a background `bash` task when asked for a persistent terminal.
+
+## Alternatives considered
+
+**Duplicate the shared rows into both overlays instead of promoting them.** Rejected on the one-home rule: three of the new rows would exist twice with no reason for the copies to diverge, and the next roster change would have to remember both.
+
+**Sandbox the TUI in the same change.** The TUI mounts unrestricted executors, which is a real gap, and an earlier revision of this change closed it by moving the confined stack into the base. Rejected as a separate decision that does not belong in a roster change: it alters what an existing surface does rather than what it offers, and it needs its own evidence — not least because the TUI has no `approval/request` answerer, so an escalation there fails closed instead of prompting.
+
+**Enable Code Mode.** Its trust posture is bash-equivalent by design and its tool calls pass the same `tools/pre-execute` gate as bash, so it is not the same call as the model-code tools above. Rejected here anyway: `both` changes every model-visible request on both surfaces, and `code` replaces the wire rather than adding to it — either is a presentation decision, not a roster one.
+
+**Mount an MCP server by default.** Rejected because a shipped default would have to name one, and any choice spawns a third-party child process on every user's machine outside the sandbox. The dependency ships instead.
+
+## Consequences
+
+The same model gets the same tools on both surfaces, and the difference that existed for no recorded reason is gone. Twenty-seven names, asserted exactly and identically on both sides, so a later change that adds a tool to one surface only — or drops one from either — fails a check instead of shipping quietly.
+
+`apps/cli` gains five workspace dependencies: four the shipped tree now mounts, plus `dsh-mcp-client`, which it does not mount and which exists so an installed `dsh` can.
+
+Nothing about execution changed. The TUI still runs the model's commands through unrestricted executors with no approval seam, and the Web surface still defaults to `danger-full-access`. Both are pinned by assertions in this change, which makes them visible rather than fixed — the sandbox decision is still open.

.agents/notes/implemented/feature/2026-07-31-even-out-shipped-tool-rosters.zh.md

@@ -0,0 +1,69 @@
+# Agent Note: 拉平交付的工具清单
+
+Status: implemented
+
+[English](2026-07-31-even-out-shipped-tool-rosters.md) | 中文
+
+## 问题
+
+两个交付的 `dsh` surface 提供着不同的工具,而没有任何记录说明为什么。会话检查点、工具结果裁剪、goal 工具和 Ralph 在 `tui.cordis.yml`;`tool-todo` 以及后来的 web 搜索在 `web.cordis.yml`。两个 surface 都没有会话搜索、字符串替换编辑器和重复工具守卫,尽管这三者都已成包存在,且没有一个是 surface 专属的。
+
+结果是一处没人做过决定的用户可见差异:同一个模型、同一个请求,在终端上能定目标而在浏览器里不能,在浏览器里能搜网页而在终端上不能。
+
+## 决策
+
+那些并非 surface 专属的行移入 [`base.cordis.yml`](../../../../apps/cli/config/base.cordis.yml),另有三行加入:`tool-session-query`、`tool-str-replace-editor` 和 `repeat-tool-guard`。Web 搜索也一并移入,这正是把它做成 Web 默认的那次改动明确推迟的 TUI 决定。两个 surface 现在组装出同样的二十七个工具。
+
+有两行仍是 surface 专属。`tmux-context` 只在 TUI,因为浏览器 surface 没有终端复用器可描述。`session-reference` 只在 TUI,因为它以 launcher 的进程本地路径驱动共享的 session-query 索引,而浏览器侧边栏会在自己的首次搜索里重建该索引。
+
+**本次改动只做加法。** 两个 surface 都没有任何一行被移除,也没有任何既有行的配置被编辑:执行器、沙箱组合、访问默认值、`tools.mode` 以及 workflow 工具,全都保持原样。对比改动前后的两份目录,读者应当只看到新增,别无其他。
+
+### 什么保持不挂,以及为什么
+
+有三项能力基于其自身包所记录的证据保持在外,列在这里是为了让「我们忘了」和「我们决定不要」保持可区分。
+
+**`dsh-tool-cordis`** 让模型写一段 JavaScript 并挂成临时插件。它的 README 写明了这个界限:「The sandbox is containment for honest code, not a security boundary — host-realm helpers on the sandbox global are reachable, so mount code can reach Node」([Known limitations](../../../../packages/cordis/tool-cordis/README.md))。`node:vm` 的 realm 就在 harness 进程内,而 `dsh-sandbox-local` 只约束它 spawn 出去的 argv,因此在 Web surface 上,沙箱与批准接缝是被绕过而非被执行。
+
+**`dsh-web-fetch-local`** 保持不挂,`dsh-tool-web` 保持 `fetch: false`。SSRF 防护在实现中是 deferred 状态([`policy.ts`](../../../../packages/web/web-fetch-local/src/policy.ts) 只校验协议、凭据与长度),包里也直说了:「this provider is an SSRF primitive and **must not be enabled** in a deployment that can reach sensitive internal network targets」([README](../../../../packages/web/web-fetch-local/README.md))。目标由模型选择,其中包括 harness 自己跑在环回地址上的网关、内网段和云元数据端点。
+
+弃用它收窄的是接触面而非可达性:`bash` 是挂着的,`curl` 照样能拿到同一个页面——一次真实运行确认了这点。这个缺席买到的是去掉一个无需 shell、以参数成形的请求原语,以及随之而来的那条意外路径:一次「帮我总结这个页面」悄悄打到环回地址。真要收住出站流量的部署需要的是网络层管控。
+
+**LSP 三件套**留在外面是运维原因而非安全原因:`command` 在插件加载时从 `PATH` 解析,因此缺少语言服务器会让整次启动失败,而不只是失去一个工具。等到「缺失」退化为「跳过注册」之后,它就可以挂了。
+
+### MCP 是依赖,不是配置行
+
+`@deepseek-ai/dsh-mcp-client` 成为本 CLI 的运行时依赖,但在任何交付配置里都没有对应的行。该插件每个实例只挂载一台服务器,且 `command` 是必填,因此一个默认值必须点名一台第三方服务器,并在每次启动时把它作为子进程 spawn——不经 `ctx.bash`,因而也在 Web surface 所组合的沙箱策略之外。
+
+真正能让 MCP 成为默认的那一层,恰恰是本仓库尚未拥有的:一个读取用户服务器清单、按条目逐台挂载客户端的桥接,形态与 [`dsh-hooks-claude`](../../../../packages/hooks/hooks-claude/README.md) 读取 Claude Code 的 `hooks.json` 完全相同。交付这个依赖意味着已安装的 `dsh` 今天就能从 `$DSH_HOME/config.yaml` 挂载服务器;CLI README 里给了那段 YAML。
+
+## 测试
+
+[`apps/cli/tests/shipped-composition.e2e.ts`](../../../../apps/cli/tests/shipped-composition.e2e.ts) 在伪终端中通过真实 Loader 启动交付树,并从会话日志持久化的 `request/header` 中读出工具名,因此断言的正是模型实际收到的目录。它传入的 `--config` overlay [`composition-keyless-tail.cordis.yml`](../../../../apps/cli/tests/fixtures/composition-keyless-tail.cordis.yml) 只做测试隔离:一个无网络适配器,以及落在工作区内的会话产物。
+
+该尾部还插入了 [`composition-settled.ts`](../../../../apps/cli/tests/fixtures/composition-settled.ts),它在终端字节流上宣告 Loader 激活已 settle。TUI 在自己的 fiber 一启动就渲染,因此在 banner 处敲下的提示词可能在工具行与持久化仍在激活时就抵达循环,从而组装出不完整的目录;把冒烟的首个提示词 gate 在该标记上,正是断言得以确定的原因。
+
+同一份冒烟还从同一份产物上钉住 TUI 未改变的执行姿态:`tool-bash` 只在挂载的执行器确实有更宽模式可升级时才发出 `sandbox_permissions` 升级参数对,因此断言它的**缺席**会在日后有人悄悄给这个 surface 加上沙箱时失败。
+
+[`apps/web/tests/shipped-composition.e2e.ts`](../../../../apps/web/tests/shipped-composition.e2e.ts) 在构建产物 lane 中覆盖 Web surface,断言它的工具目录、它的访问默认值未被触碰,以及 `workspace-write` 的可写根包含临时目录——一个会让沙箱测试说谎的陷阱,当工作区落在 `/tmp` 下时([`roots.ts`](../../../../packages/sandbox/sandbox/src/roots.ts))。
+
+`glob` 与 `grep` 被作为全有或全无的一对断言,而不是固定成员:`dsh-tool-fs-search` 在加载时探测 `command -v rg`,没有 ripgrep 就两个工具都不注册,这是宿主依赖。
+
+除入库测试外,两个 surface 都以 plain Node 从构建产物 `apps/cli/lib/bin.js` 出发、用真实密钥驱动过。每一个已挂载的工具都执行成功,包括 `ralph` 与 `web_search`;模型从未触达 `cordis_*` 或 `mcp_*`,被要求做 LSP 跳转时退化到 `grep`,被要求开持久终端时用了后台 `bash` 任务。
+
+## 曾考虑的替代方案
+
+**把共享的行复制进两份 overlay,而不是提升到 base。** 基于「一处归属」原则否决:新增行里有三行会存在两份,而这些副本没有任何理由发生分歧,下一次改工具清单还得记着改两处。
+
+**在同一次改动里给 TUI 加沙箱。** TUI 挂的是不受限执行器,这确实是个真实缺口,本次改动的早先一版曾把受限栈移入 base 来堵上它。作为一个不属于工具清单改动的独立决定被否决:它改变的是一个既有 surface 的行为而非它提供的东西,而且需要它自己的证据——尤其因为 TUI 没有 `approval/request` 的应答方,升级请求在那里是 fail-closed 而不是弹出提示。
+
+**开启 Code Mode。** 它的信任立场按设计与 bash 同级,工具调用要过与 bash 相同的 `tools/pre-execute` 闸门,所以它与上面那些模型写码工具不是同一个判断。在这里仍被否决:`both` 会改变两个 surface 上每一个模型可见请求,而 `code` 是把线路替换而非加一个——两者都是呈现方式的决定,不是工具清单的决定。
+
+**默认挂一台 MCP 服务器。** 否决,因为交付默认值必须点名一台,而任何选择都会在每个用户的机器上、在沙箱之外 spawn 一个第三方子进程。改为交付依赖。
+
+## 后果
+
+同一个模型在两个 surface 上拿到同样的工具,那处没有记录理由的差异消失了。二十七个名字,两侧精确且一致地断言,因此日后只给一个 surface 加工具、或从任一侧丢掉工具,都会让检查失败而不是悄悄发出去。
+
+`apps/cli` 增加五个 workspace 依赖:四个是交付树现在挂载的,外加 `dsh-mcp-client`——它并不被挂载,存在的意义是让已安装的 `dsh` 能挂。
+
+执行相关的一切都没有变。TUI 仍以不受限执行器运行模型的命令且没有批准接缝,Web surface 仍默认 `danger-full-access`。两者都由本次改动中的断言钉住,这让它们变得可见而非被修复——沙箱那个决定仍然悬着。