dsh-plugin-translate-flow

dsh-plugin-translate-flow

Isolated, customizable translation flow for DeepSeek Harness.

The plugin exposes one page:

• 翻译流程 / Translation Flow — input node, draggable module cards, an output node after every module, and sequential execution with progress highlighting. Each module's prompt and model are edited in place, from that module's card.

Built-in modules: 中译英, 英译中, 学术中文润色, 学术英文润色. The default mode is the precision chain:

EN input -> 英译中 -> 学术中文润色 -> 中译英

Every module is a separate model call carrying only its own prompt and its own input; execution is strictly isolated.

Presets

Two buttons in the page header replace the chain with a precision round trip:

• 汉语精译 / Chinese precision — zh2en -> en-academic -> en2zh
• 英文精译 / English precision — en2zh -> zh-academic -> zh2en

Each one translates out of the source language, polishes in the target language and translates back, so what the polish did is visible. 英文精译 is also the chain a fresh panel starts from, and the chain the agent tool runs when steps is omitted. A preset is a starting point rather than a mode: the modules it places are ordinary modules, editable and removable like any other.

English name

Published community display name: Translation Flow.

Isolation, measured rather than asserted

Each module call is assembled in one place (runIsolatedModule in lib/index.js) and carries:

• system — the module's own prompt, with nothing concatenated onto it;
• messages — a fresh one-element history holding exactly this module's input;
• no tools, no toolHistory, no sessionId, no purpose;
• a frozen options object, so an llm/stream listener cannot rewrite it in flight.

The isolation badges in the viewer are read off the object handed to llm.stream, so the numbers follow the request rather than restating a claim.

Model routing

Three tiers, in order:

1. the module's own provider/model override — editable on each module's card;
2. the loader row config: { provider, model } — no panel control: set it in the loader row if you want a durable default that is not the session's model;
3. the session's default agent model.

OpenCode-style providers requiring a per-conversation x-opencode-session header are not supported by the deliberately session-less calls; use a plain Messages/OpenAI-compatible provider.

Per-call timeout

The default budget is 300 s per module call. A loader row may override it:

- id: translate-flow
  config:
    provider: <provider>
    model: <model>
    timeoutMs: 120000   # 1 s .. 10 min; anything else is ignored

The budget is enforced twice: the abort signal asks the provider to stop, and a deadline inside consumeStream stops waiting. A provider that never yields and ignores the signal is therefore reported as a timeout instead of holding the run and reserving its runId forever, and cancelling such a run resolves at once.

Agent tool

The same isolated chain is reachable by the agent as the translate_flow tool: one call per step, no session attached, and omitting steps runs the built-in precision chain. It is registered only when the host exposes a tool registry, so a host without one still activates the panels.

Token usage

When a provider reports token accounting, each module's result carries it and the viewer's info tab shows prompt/input sizes, token counts and elapsed time beside the isolation record.

Compatibility

• DeepSeek Harness 0.2.0-rc.2 (desktop) tested.
• Node.js >=18.
• No extra runtime npm dependencies; React is provided by the host.

Install

Install as a bundle (the package declares its own patch, which inserts its loader row):

dsh plugin --profile <name> add dsh-plugin-translate-flow

For an application-owned profile (such as the Desktop one), install in-app with the plugin manager, passing the absolute package directory as the target.

The page then appears as the sidebar entry 翻译流程 / Translation Flow, addressing its own main panel.

Host-half changes are cached by package name for the life of the process (both Node's ESM module cache and the Typert manifest cache), so replacing an installed host half needs a DSH restart or a new package name.

Privacy and accounting

Because the calls carry no sessionId, they are not part of the session log:

• they do not appear in the session transcript;
• they are not attributable to a session in token-usage accounting;
• they are not visible in per-session dashboards.

This is a direct consequence of the isolation design — a real trade-off, not an oversight. To keep auditable usage, do not use this plugin for it. The text you translate does go to whichever provider you selected, exactly as it would from the agent itself.

Verification

node test/host.test.mjs      # 120 assertions
node test/manifest.test.mjs  # 101 assertions
node test/client.test.mjs    #  56 assertions

277 assertions covering the isolation contract, the two presets and the flows they build, the graph-repair edge cases a hand-edited or imported document can produce, the declared-shape check on the agent tool's own result, deadline and cancellation behaviour, the token-usage path, the identifiers shared by the three faces, and a one-pass render of the page.

Publishing checklist

• dsh.bundle.patch points at a file that exists and parses
• cordis.patch.yml names the package in its loader row
• files lists everything the runtime needs
• all three test files pass
• README / README.zh.md / LICENSE / icon / locale are current

Community topics: dsh-plugin, deepseek-harness, dsh, cordis.

License

MIT — see LICENSE.