dsh-chinese-language
Make a DeepSeek Harness (dsh) agent think and answer in Simplified Chinese — through a token-lean system-prompt section plus a one-line runtime-context reminder, without touching the deployment persona.
English | 中文
Why
DeepSeek Harness ships an English-first system prompt. The model's hidden reasoning follows that prompt, and the visible answer follows the reasoning, so unless something explicitly says otherwise a long session drifts into English.
Telling the agent "answer in Chinese" in every message works, but it is per-session, easy to forget, and gone the moment you resume in a new window. This plugin states the rule where the model always reads it.
What it does
The plugin registers one rule through two channels with different token profiles:
| Channel | Where it lands | Cost |
|---|
System-prompt section user:chinese-language, order 10300 | after deployment:persona-suffix (order 10200), i.e. the very end of the assembled system prompt | Resent with every request; kept short for that reason |
Runtime-context reminder user:chinese-language, order 100 | the dynamic-context snapshot the harness projects next to the newest messages | Persisted once and retained: the harness writes a new snapshot only when the text changes, and this text is static |
Why both: a fresh session is covered by the section alone. A long or resumed session carries a history full of English reasoning, where a single rule at the head of that history loses influence — the runtime-context reminder restates it next to the live conversation, which is what makes an already-running session pick the rule up without being restarted.
Both registrations are purely additive: the plugin never replaces personaPrefix / personaSuffix, never rewrites a tool, skill, or workspace-instruction row, and never reorders a first-party section. It covers every agent that inherits the mounting scope, so a root mount reaches every session and every subagent that joins it.
The default section is two lines, and the reminder is one shorter line:
语言规则：始终使用简体中文进行内部推理（thinking）与最终回答，历史内容为英文时同样适用。
代码、命令、路径、URL、报错原文和专有名词保留原文。
语言提醒：本轮内部推理与可见回答使用简体中文。
Token budget
| Version | Steady-state default text |
|---|
| 0.1.0 (section only, four lines) | ~148 characters |
| 0.2.0 (two lines + one reminder line) | ~99 characters |
Two properties keep the cost flat:
- The section text is short, and it sits at a stable position inside the prompt prefix, so it does not invalidate the prefix cache.
- The runtime-context text never changes, so the harness persists it once and keeps reusing that message. It does not accumulate per step, and it adds no per-turn reminder text.
Turn the reminder off entirely with context: false if you want the absolute minimum; on the 0.1.7 line the old behavior is exactly context: false plus a four-line lines list.
Install
1. From GitHub (recommended)
dsh plugin --profile web add github:fzy3232/dsh-chinese-language
dsh web # restart to apply, then start a new session
2. From npm, once published
dsh plugin --profile web add dsh-chinese-language
Either channel works the same way: the package declares dsh.bundle.patch, so dsh plugin add appends dsh-chinese-language to dsh.profile.bundles, and boot merges the bundled cordis.patch.yml, which inserts the plugin row. No profile file is edited by hand.
Replace web with your profile name (tui, headless, ...) if you are not on the Web profile.
3. Single-file local mount (no package manager)
The plugin is one dependency-free ES module, so it also works as a local file.
mkdir -p "$DSH_HOME/dsh-plugins"
curl -fsSL -o "$DSH_HOME/dsh-plugins/chinese-language.mjs" \
https://raw.githubusercontent.com/fzy3232/dsh-chinese-language/main/lib/index.js
Then add one insert to $DSH_HOME/cordis.patch.yml (create the file if it does not exist):
- insert:
- id: user-chinese-language
name: ./dsh-plugins/chinese-language.mjs
Upgrading
A GitHub dependency is pinned to the installed commit, so a running profile keeps the old build until you refresh it:
dsh plugin --profile web update dsh-chinese-language
dsh web # restart the profile to load the new module
Already-running sessions keep the system prompt they were assembled with; the runtime-context reminder reaches them on their next step after the restart.
Configuration
Every field is optional. Unknown keys are ignored, and an unusable value falls back to that key's default instead of failing the loader entry.
| Key | Default | Meaning |
|---|
lines | the two built-in rules | Section lines, joined with newlines. Empty or non-string entries are dropped. |
text | - | Replaces lines entirely when non-empty. Wins over lines. |
order | derived, else 10300 | Section sort order. Anything above 10200 lands after the first-party sections. |
section | enabled, named user:chinese-language | Either false to skip the section channel, or a string naming it. |
sectionName | user:chinese-language | Section name. Must stay unique within the scope; a duplicate name throws. |
contextText | the one-line reminder | Replaces the reminder text without touching the section text. |
context | enabled, named user:chinese-language | Either false to skip the reminder, or a string naming it. |
contextName | user:chinese-language | Reminder name, unique within the scope. |
contextOrder | derived, else 100 | Reminder sort order; lower values print earlier in the snapshot. |
On harnesses that expose getSectionOrder() / getContextOrder(), the two orders are derived from DEPLOYMENT_PERSONA_SUFFIX and SANDBOX_POLICY, so they still track the registry if those slots move. Older harnesses use the constants.
Example, keeping both channels but tightening the wording:
- insert:
- id: dsh-chinese-language
name: dsh-chinese-language
config:
lines:
- 始终使用简体中文回答。
- 代码、命令和报错原文保留原文。
contextText: 提醒：用简体中文。
A patch replaces the targeted row's whole config, so restate every key the row needs whenever an override adds one.
Verify
dsh --profile web --dump-config | grep -A2 dsh-chinese-language
For a local mount, look for your file path instead:
# == /Users/you/.dsh/cordis.patch.yml
- id: user-chinese-language
name: file:///Users/you/.dsh/dsh-plugins/chinese-language.mjs
Then start a new session: the rule reaches the model from the first turn. In a session that is already running, the reminder arrives on its next step — the runtime-context snapshot the harness shows above the newest message carries a contribution named user:chinese-language:
Current runtime context. This snapshot supersedes earlier runtime-context snapshots.
...
语言提醒：本轮内部推理与可见回答使用简体中文。
That snapshot is written once and then replaced in place, so the reminder neither accumulates nor repeats: only the changing parts of the snapshot (sandbox state, recalled memory, ...) cause a new one.
When a channel is unavailable
The two channels degrade independently. The rule never disappears because one of them is missing:
| Situation | Section channel | Reminder channel |
|---|
| Normal | yes | yes |
Harness without systemPrompt.context() (older than the runtime-context API) | yes | skipped |
Deployment sets includeRuntimeContext: false on dsh-system-prompt | yes | suppressed by the harness |
An agent preset claims the prompt with a complete: true persona section | replaced — the harness restores that section as the sole prompt section | yes |
The last row is the reason the reminder exists rather than being a duplicate: a complete persona is exactly the case where a section-only language plugin goes silent. Registering one rule through both channels keeps at least one of them in front of the model.
Both channels were exercised against every @deepseek-ai/dsh-system-prompt release installed on the development machine — 0.1.5-rc.2, 0.1.5-rc.3, 0.1.7-rc.1 and 0.1.7-rc.2 — and mount identically on each, with the section staying last. Releases older than 0.1.5-rc.2 are untested; they keep the section channel and skip the reminder.
Scope and limits
- This constrains the prompt — the system section plus the runtime-context snapshot — which is what the model reads before it starts reasoning. It is not a token-level filter.
- Hidden reasoning is model output. A prompt strongly steers it, but nothing at this layer can give a mathematical guarantee; expect the rule to hold in practice rather than absolutely.
- Code, commands, file paths, URLs, API names, and verbatim error text are meant to stay in their original form. That is part of the shipped rule.
- If another plugin registers a section or a context with the same name in the same scope, the loader fails loudly. Change
sectionName / contextName to resolve it.
Uninstall
dsh plugin --profile web remove dsh-chinese-language
For a local mount, delete the insert entry from $DSH_HOME/cordis.patch.yml and remove the .mjs file.
Development
There is no build step and no runtime dependency. lib/index.js is the whole plugin.
npm test # node --test: contract + manifest (+ real-harness when available)
node --check lib/index.js
| Suite | What it covers |
|---|
test/contract.test.mjs | Both channels, every config key, the older-harness path, the steady-state token budget |
test/manifest.test.mjs | npm pack --dry-run carries the module, the patch, the READMEs and the license |
test/rc2-integration.test.mjs | Assembles against the real @deepseek-ai/dsh-system-prompt: section last, reminder in the snapshot, complete persona, suppressed runtime context, and every release found in the profile's pnpm store |
The third suite imports @deepseek-ai/*, which lives in a dsh profile rather than here, so it reports as skipped unless those packages resolve. To run it for real, point the checkout at an installed profile (the link is ignored by git):
ln -sfn "$DSH_HOME/profiles/web/node_modules" node_modules
npm test
To exercise it against a real tree, mount it with the local-mount recipe above and re-run dsh --profile web --dump-config.
License
MIT. See LICENSE.
