dsh-encoding

Correct-encoding command runner and mojibake recovery for DeepSeek Harness.

让 DeepSeek Harness 在 Windows 上不再乱码。

English / 中文

---

English

dsh-encoding is a thin tools plugin for DeepSeek Harness (dsh). On Chinese Windows, legacy CLIs, Git Bash/MSYS2 and PowerShell 5.1 often write GBK (cp936) or UTF-16LE bytes, which dsh's built-in bash/pwsh tools decode as UTF-8 and show as mojibake. The decoding happens inside the non-pluggable core service (dsh-subprocess-local), so this plugin does not try to patch it. Instead it adds two complementary tools:

Tool	What it does
encoding_run	Spawns its own child process, captures raw bytes of stdout/stderr, auto-detects UTF-8 / UTF-16LE / GBK and decodes them correctly.
encoding_decode	Recovers already-garbled text (e.g. GBK bytes misread as latin1) back to correct Chinese.

0.2.0 breaking change: tool names were renamed from run/decode_text to encoding_run/encoding_decode (collision-safe prefix). 0.1.0 is deprecated on npm.

Relation to official built-ins (verified on dsh 0.2.0-rc.2)

• dsh 0.2.0 does not ship built-in run / decode_text tools. Verified with a clean, plugin-free profile (dsh --profile pmaudit --dump-config): the base config contains no such tool rows (built-in tools are tool-bash, tool-pwsh, tool-fs, …), and the official tools reference documents none.
• The built-in bash/pwsh tools still have no encoding configuration — the shell seam's ShellExecRequest has no encoding field — and a GBK-producing command through the built-in pwsh tool still returns mojibake on this environment. This plugin's premise therefore holds.
• Tool names are prefixed so they cannot collide with built-in or future dsh tools: the registry rejects duplicates within one layer and the reserved run_code name (official tools.md).

Install

From npm:

dsh plugin --profile <name> add dsh-encoding

From a local checkout (this repo):

dsh plugin --profile <name> add ./dsh-encoding

From GitHub (works with no build authorization because the package is pure ESM JavaScript with no build step):

dsh plugin --profile <name> add github:Lostforest7/dsh-encoding

Verify the layer is active, then start:

dsh --profile <name> --dump-config   # shows a "# == dsh-encoding" layer
dsh --profile <name>

Tool: encoding_run

Parameter	Type	Required	Default	Description
command	string	yes	—	Command line to execute (Windows: cmd /d /s /c, POSIX: /bin/bash -c).
workdir	string	no	—	Working directory for the child process.
timeoutMs	number	no	120000	Timeout; the whole process tree is killed on expiry.
encoding	string	no	auto	auto (detect) or force utf-8 / utf-16le / gbk.

Example output (GBK output on Windows):

中文
[stderr]
... (if any stderr)
[encoding: gbk/utf-8 (auto-detected)]
[exit code: 0]

The last line is always [exit code: N]. When the output contains U+FFFD (�), an explicit warning line is appended instead of silently pretending the decode was right.

Positioning, stated plainly: encoding_run is an independent new tool. It does not have the built-in bash tool's sandbox, approvals, or background-task support, and it executes with the same permissions as the harness process.

Tool: encoding_decode

Parameter	Type	Required	Default	Description
text	string	yes	—	The garbled text to recover.
from	string	no	auto	How to turn characters back into bytes: latin1 / utf-16le / auto.
to	string	no	auto	Target decoding: gbk / utf-16le / auto.

The classic mojibake path is from=latin1, to=gbk: GBK bytes were misread as latin1 (one character per byte), so the tool re-encodes each character to one byte and decodes the bytes as GBK.

With auto, the direction is guessed: embedded NULs → latin1 → utf-16le; wide characters → utf-16le → gbk; everything else → the classic latin1 → gbk.

Irreversible cases are reported, never faked: if the input already contains U+FFFD (�) — i.e. the shell already destroyed the bytes with an earlier mis-decode — the result is ok: false with an explicit note.

Note: from=gbk is rejected with a clear explanation. Node.js ships a GBK decoder (TextDecoder) but no GBK encoder, and this plugin is zero-dependency, so re-encoding characters back to GBK bytes is not possible. to=gbk is fully supported.

30-second example

Built-in bash shows ÖÐÎÄ where 中文 was expected (GBK bytes read as latin1):

1. Ask the agent to use encoding_decode with {"text": "ÖÐÎÄ"} → 中文 — done.
2. Or re-run the command with correct decoding: use encoding_run with {"command": "chcp 936 >nul & echo 中文"} → output 中文 plus [encoding: gbk (auto-detected)] [exit code: 0].

CLI smoke test (no DSH required)

The CLI subcommands keep their short names (run / decode_text) — they are a dev smoke runner, not the registered tool names:

node src/cli.js decode_text "ÖÐÎÄ"                  # -> 中文
node src/cli.js run "chcp 936 >nul & echo 中文"      # GBK output, auto-decoded
node src/cli.js run "echo hello"
node src/cli.js detect D6D0CEC4                      # -> detected: gbk / decoded: 中文

Development

Zero dependencies — none at runtime, none for development; tests use only Node built-ins (node:test, node:assert). No npm install, no DSH, no build step needed:

node --test test/

Architecture:

dsh-encoding/
├── package.json          # type: module, dsh.bundle.patch -> cordis.patch.yml
├── cordis.patch.yml      # inserts one loader row: id "encoding", name "dsh-encoding"
├── src/index.js          # Cordis entry (dependency-free), registers encoding_run + encoding_decode
├── src/encoding.js       # pure functions: detect / decode / mojibake recovery
├── src/exec.js           # spawn wrapper + run formatting (spawnImpl injectable)
├── src/cli.js            # manual smoke runner (no DSH required)
└── test/                 # node:test suites, mock child_process.spawn

src/encoding.js and src/exec.js import nothing from @deepseek-ai/*. src/index.js is dependency-free too: it emits the registry-ready tool shape directly (the same {type:'object', properties, required} + output.schema/render form that the official defineTool compiles to — locked by literal-schema assertions in test/index.test.js, and accepted by the live 0.2.0-rc.2 registry). This is deliberate — when a plugin is installed as a local checkout (dsh plugin add ./path, a pnpm link), the package lives outside the profile tree, and Node's ESM resolver cannot reach the host's @deepseek-ai/* packages from the checkout directory (the import dies with ERR_MODULE_NOT_FOUND, verified on 0.2.0-rc.2). Emitting the registry shape directly keeps the plugin working under npm, git, and local-link installs alike.

Known limitations

• Does not replace or fix the built-in bash/pwsh tools, ctx.shell, or dsh-subprocess-local — those are out of scope by design (version-fragile core work). This plugin is a supplement, and says so.
• No sandbox / approval / background-task integration for encoding_run.
• Encoding detection is heuristic (UTF-8 validity, UTF-16LE NUL pattern, GBK high-byte ratio, BOM). Ambiguous cases exist — e.g. BOM-less pure-CJK UTF-16LE vs GBK — so force encoding when you know the answer.
• U+FFFD in the input is information already destroyed; recovery is impossible and is reported as such.
• Output is capped at 8 MiB per stream; on timeout the process tree is killed (Windows: taskkill /T /F, POSIX: process-group SIGKILL).

Publishing checklist

• GitHub repository with topic dsh-plugin — Lostforest7/dsh-encoding
• npm name dsh-encoding (was unoccupied; published 0.1.0 → 0.2.0)
• npm publish --registry=https://registry.npmjs.org (China mirrors are read-only; always publish against the official registry)
• Submit to the awesome-dsh-plugin list and the dsh-plugin.org market (optional, not yet done)

中文

dsh-encoding 是 DeepSeek Harness（dsh）的一个薄工具插件。在中文 Windows 上，老式 CLI、Git Bash/MSYS2、PowerShell 5.1 经常输出 GBK（cp936）或 UTF-16LE 字节，被 dsh 内置的 bash/pwsh 工具按 UTF-8 解码后变成乱码；而解码发生在不可插拔的核心服务（dsh-subprocess-local）里。本插件不去修补内置 bash，而是新增两个补充工具：

工具	作用
encoding_run	自己 spawn 子进程、按原始字节收 stdout/stderr，自动识别 UTF-8 / UTF-16LE / GBK 并正确解码。
encoding_decode	把已经乱码的文本（如 GBK 字节被当 latin1 误读）还原成正确中文。

0.2.0 破坏性变更：工具名由 run/decode_text 改为 encoding_run/encoding_decode（防撞名前缀）。0.1.0 已在 npm 标记废弃。

与官方内置工具的关系（在 dsh 0.2.0-rc.2 上实测）

• dsh 0.2.0 没有内置 run / decode_text 工具。用干净的、未装插件的临时 profile（dsh --profile pmaudit --dump-config）验证：base 配置中没有任何同名工具行（内置工具是 tool-bash、tool-pwsh、tool-fs 等），官方 tools 参考文档也没有这两个工具。
• 内置 bash/pwsh 依然没有编码配置（shell 接口的 ShellExecRequest 无编码字段）；在本环境用内置 pwsh 工具跑 GBK 输出命令，结果仍是乱码。本插件的立论前提成立。
• 工具名加了 encoding_ 前缀，避免与内置或未来 dsh 工具撞名——注册表会拒绝同层重复名与保留名 run_code（官方 tools.md）。

安装

从 npm：

dsh plugin --profile <name> add dsh-encoding

从本地目录（本仓库）：

dsh plugin --profile <name> add ./dsh-encoding

从 GitHub（纯 ESM JavaScript、无构建步骤，不需要 prepare 构建授权）：

dsh plugin --profile <name> add github:Lostforest7/dsh-encoding

验证配置层已生效，然后启动：

dsh --profile <name> --dump-config   # 能看到 "# == dsh-encoding" 层
dsh --profile <name>

工具 encoding_run

参数	类型	必填	默认	说明
command	string	是	—	要执行的命令（Windows：cmd /d /s /c；POSIX：/bin/bash -c）。
workdir	string	否	—	子进程工作目录。
timeoutMs	number	否	120000	超时毫秒数，到期后杀掉整个进程树。
encoding	string	否	auto	auto 自动检测，或强制 utf-8 / utf-16le / gbk。

输出示例（Windows 上的 GBK 输出）：

中文
[stderr]
...（如有 stderr）
[encoding: gbk/utf-8 (auto-detected)]
[exit code: 0]

最后一行固定是 [exit code: N]。当输出含 U+FFFD（�）时，会附加明确的警告行，而不是假装解码正确。

定位如实说明：encoding_run 是一个独立的新工具，不带内置 bash 工具的沙箱、审批与后台任务能力，并以 dsh 进程自身的权限执行命令。

工具 encoding_decode

参数	类型	必填	默认	说明
text	string	是	—	要还原的乱码文本。
from	string	否	auto	字符还原成字节的方式：latin1 / utf-16le / auto。
to	string	否	auto	目标解码：gbk / utf-16le / auto。

经典"乱码还原"路径是 from=latin1, to=gbk：GBK 字节被当 latin1 误读（一个字符一个字节），工具把每个字符还原成一个字节，再按 GBK 解码。

auto 模式下自动判断误读方向：文本含 NUL → latin1 → utf-16le；含宽字符 → utf-16le → gbk；其余 → 经典 latin1 → gbk。

不可逆情况明确说明、不伪造结果：若输入里已经含 U+FFFD（�）——即字节早已在上游误解码中被毁掉——返回 ok: false 并附说明。

注意：from=gbk 会被明确拒绝并解释原因——Node.js 内置了 GBK 解码器（TextDecoder）但没有 GBK 编码器，本插件又坚持零依赖，因此无法把字符重新编码回 GBK 字节。to=gbk 完全支持。

30 秒示例

内置 bash 把 中文 显示成了 ÖÐÎÄ（GBK 字节被当 latin1 读）：

1. 让 agent 调用 encoding_decode，参数 {"text": "ÖÐÎÄ"} → 得到 中文，完成。
2. 或者重新执行命令并按正确编码解码：调用 encoding_run，参数 {"command": "chcp 936 >nul & echo 中文"} → 输出 中文，并附 [encoding: gbk (auto-detected)] [exit code: 0]。

CLI 手动冒烟（无需 DSH）

CLI 子命令保留短名（run / decode_text）——它们只是开发冒烟入口，不是注册的工具名：

node src/cli.js decode_text "ÖÐÎÄ"                  # -> 中文
node src/cli.js run "chcp 936 >nul & echo 中文"      # GBK 输出自动解码
node src/cli.js run "echo hello"
node src/cli.js detect D6D0CEC4                      # -> detected: gbk / decoded: 中文

本地测试

零依赖——运行时与开发时都没有任何依赖；测试只用 Node 内置的 node:test / node:assert，无需 npm install、无需 DSH、无需构建：

node --test test/

结构：

dsh-encoding/
├── package.json          # type: module，dsh.bundle.patch -> cordis.patch.yml
├── cordis.patch.yml      # 插入一行 loader：id "encoding"，name "dsh-encoding"
├── src/index.js          # Cordis 入口（零依赖），注册 encoding_run 与 encoding_decode
├── src/encoding.js       # 纯函数：检测 / 解码 / 乱码还原
├── src/exec.js           # spawn 封装 + run 输出组装（spawnImpl 可注入）
├── src/cli.js            # 手动冒烟（无需 DSH）
└── test/                 # node:test 套件，mock child_process.spawn

src/encoding.js 与 src/exec.js 不 import 任何 @deepseek-ai/*；src/index.js 同样零依赖：它直接输出注册表形态的工具定义（与官方 defineTool 编译产物一致的 {type:'object', properties, required} + output.schema/render，由 test/index.test.js 的字面 schema 断言锁定，并已被 0.2.0-rc.2 真实注册表接受）。这是刻意的——当插件以本地目录方式安装（dsh plugin add ./path，即 pnpm link）时，包位于 profile 树之外，Node 的 ESM 解析无法从插件目录找到宿主的 @deepseek-ai/* 包（会以 ERR_MODULE_NOT_FOUND 失败，已在 0.2.0-rc.2 实测）。直接输出注册形态让插件在 npm、git、本地 link 三种安装方式下都能工作。

已知限制

• 不替换、不修复内置 bash/pwsh、ctx.shell、dsh-subprocess-local——刻意超出范围（核心层改动版本脆弱）。本插件定位为"补充工具"。
• encoding_run 无沙箱集成、无审批、无后台任务。
• 编码检测是启发式的（UTF-8 合法性、UTF-16LE NUL 模式、GBK 高位字节占比、BOM）。存在歧义场景——例如无 BOM 的纯中文 UTF-16LE 与 GBK 难以区分——已知编码时请用 encoding 参数强制指定。
• 输入中已有的 U+FFFD 属于已丢失的信息，无法还原，并会如实报告。
• 每路输出上限 8 MiB；超时杀进程树（Windows：taskkill /T /F；POSIX：进程组 SIGKILL）。

发布清单

• GitHub 仓库已建并带 topic dsh-plugin：Lostforest7/dsh-encoding
• npm 名 dsh-encoding（确认未占用；已发布 0.1.0 → 0.2.0）
• npm publish --registry=https://registry.npmjs.org（国内镜像只读，务必用官方 registry 发布）
• 提交到 awesome-dsh-plugin 列表与 dsh-plugin.org 市场（可选，尚未完成）

License

MIT — see LICENSE.