honlnk/dsh-input-assist

English | 中文

An input assistant plugin for DeepSeek Harness (dsh):

• Inline completion (ghost text) — after a typing pause, calls the DeepSeek FIM endpoint (/beta/completions) and renders the suggestion inline right after your cursor as gray text (the Copilot look: transparent mirror layer, no floating panel). Suggestions appear progressively via SSE streaming — the first token hits the screen immediately and the ghost text grows as it generates. Tab accepts one word at a time (works mid-stream without breaking the stream), Shift+Tab accepts everything, Esc dismisses.
• Typo checking — two layers: a dictionary layer that runs entirely in your browser (228 Chinese wrong phrases + 211 English misspellings / proper-noun capitalizations + 8 context rules; the word lists live as plain text under data/ and can be merged with a per-browser custom dictionary; highlights in 200 ms, offline, zero cost) plus an LLM context pass (mixed Chinese/English: 在/再, 的/得/地, English spelling; debounced 800 ms, same cadence as completion). Errors are underlined in red inside the text (current item highlighted, click red text to select), a navigation bar walks through fixes one by one — never a bulk replace.

Completion shortcuts (active while a ghost suggestion is visible)

Action	Shortcut
Accept one word (Chinese segmented by word)	Tab
Accept all	Shift+Tab
Dismiss suggestion	Esc

The shortcut hint shows on a dock line outside the input box (same spot as the typo navigation bar) — no overlays inside the input itself.

Typo-check shortcuts (active while the input is focused and errors are found)

Action	Shortcut
Next / previous error	Ctrl+Shift+. / Ctrl+Shift+,
Fix the selected error	Ctrl+Shift+F
Mark as correct input (no replace)	Ctrl+Shift+G
Stop reminding for this draft	Esc

Quick start

dsh plugin --profile web add dsh-input-assist   # or @honlnk/dsh-input-assist — identical content
dsh --profile web          # start, then open http://127.0.0.1:3080

Local development (install from source):

git clone https://github.com/honlnk/dsh-input-assist && cd dsh-input-assist
npm install && npm run build
dsh plugin --profile web add "$PWD"

Website (demo + install guide): https://honlnk.github.io/dsh-input-assist/

API key — pick one of three (needed by completion and the LLM layer; the dictionary layer needs no configuration at all):

1. Do nothing — the plugin automatically reuses the DEEPSEEK_API_KEY already saved in dsh (.credentials.yaml) or from the environment
2. Fill it in the settings page: gear icon at the bottom of the sidebar → Plugins → Plugin Settings → Input Assist card (recommended — every option is graphical)
3. Write input-assist.completionApiKey: sk-… in ~/.dsh/settings.yaml, or run DEEPSEEK_API_KEY=sk-… dsh web

Settings-page card

Every option can be edited graphically in the dsh settings UI: gear icon at the bottom of the sidebar → Plugins → Plugin Settings, expand the “Input Assist” card. The interaction mirrors the official sibling cards (Bash / Agent Loop / Web Search): staged edits → one unified “Save / Discard”, numeric and required-text validation, per-field “unsaved” dots and per-field undo (×). The card registers through the official settings.plugin.item slot (keyed by settings namespace) — no extra services required, and external edits to the settings file hot-sync back into the card.

The model fields support fetching the catalog from the API: click the ▾ square button next to the field to open a custom dropdown that pulls the latest catalog every time it opens; click to fill. The field always stays a free-text input — any custom model name on an OpenAI-compatible endpoint can be typed by hand. If the fetch fails (no key configured, endpoint unreachable) it falls back to manual entry. Default models have followed the official migration: deepseek-chat was retired on 2026-07-24; the default is now deepseek-flash (V4.1 Flash).

Custom dictionary (per-browser only)

The “Custom dictionary (this browser only)” section at the bottom of the settings card is a paired-entry list: wrong word on the left, correct word on the right, Chinese or English. “＋ Add entry” appends a pair, × deletes a row, drag the row handle to reorder (or focus and move with ↑/↓); add / delete / reorder all animate smoothly.

• Entries are merged with the built-in dictionary for the local dictionary-layer scan: an entry with the same wrong word overrides the built-in one; mapping a word to itself (self-mapping) disables checking for that built-in entry; English entries automatically use word-boundary matching (camelCase, snake_case and dot-joined identifiers never false-positive)
• Stored only in this browser's localStorage (key dsh-input-assist:user-dict:v1, format is still one wrong => right per line) — never synced via settings.yaml
• The header row has “Import / Export / Copy”: import picks a .txt file (one wrong => right per line, same format as the built-in lists, # comments); same-word entries update in place, the rest append to the staged list — still reviewable before saving. Export downloads the current list (including unsaved edits) as the same .txt format; copy goes to the clipboard. That's how you move dictionaries between browsers and machines
• Limits: 2000 entries, wrong word ≤ 16 chars; validation issues (empty words, runs of whitespace, overlong, duplicates) show inline per row and block saving; edits share the card-level “Discard / Save” buttons, and a successful save immediately re-highlights the current draft with the new dictionary

Configuration (settings.yaml → input-assist, or edit graphically via the settings card)

Key	Default	Description
completionEnabled	true	Completion switch (same as the 「补」 button next to the input)
completionBaseUrl	https://api.deepseek.com/beta	FIM endpoint (swap for any OpenAI-compatible one)
completionApiKey	''	Empty = credential fallback
completionModel	deepseek-flash	Completion model (pick from the dropdown in the settings card, ⟳ fetches the catalog from the API)
completionDebounceMs	800	How long after typing stops before requesting (same cadence as the LLM proofread)
completionMaxTokens	64	Suggestion length cap
completionStream	true	Stream completion via SSE for progressive rendering; turn off to receive whole suggestions if the endpoint lacks streaming
proofreadEnabled	true	Typo-check switch (same as the 「校」 button)
proofreadUseLlm	true	Stack the LLM layer on top (off = zero cost)
proofreadModel	deepseek-flash	LLM proofread model (same dropdown)
proofreadDebounceMs	800	LLM-layer debounce (same cadence as completion; both return together)
proofreadDictDebounceMs	200	Dictionary-layer debounce (browser-local scan, zero cost)

Development & testing

Sources live in src/*.ts (TypeScript strict mode), built with tsdown into two entries:

• src/index.ts → lib/index.js (ESM + .d.ts, the host half, imported directly by the cordis loader)
• src/client.ts → lib/client.js (CJS, wrapped automatically by window.__ModuleLoader__.load via banner/footer, react / dsh-client-runtime kept external, local modules like the dictionary bundled in)

npm install
npm run build   # tsdown dual-entry build into lib/ (no strict tsc check)
npm test        # build → tsc --noEmit → node --test (type errors caught locally)

Dictionary data is maintained as plain text under data/ (zh-wrong-phrases.txt + en-wrong-spellings.txt + en-proper-nouns.txt line-based lists + zh-context-rules.json context rules). Before building, scripts/gen-dict.mjs parses, validates and generates src/proofread-dict-data.generated.ts (committed to git; CI guards sync via gendict + git diff) — to update the dictionaries, just edit the data/ files; npm run build picks it up. test/client-bundle.test.js executes the real bundle artifact with a fake ModuleLoader, covering the full data → bundle chain. User dictionaries (localStorage) are merged with the built-in lists for scanning, browser-side.

Architecture at a glance

Browser half (lib/client.js, bundled from src/client.ts)      Host half (lib/index.js)
  input.overlay   error hints only (no popup)       loopback    settings namespace input-assist
  input.dock      typo panel + fixes              ───RPC (0.1.5+: fetch route /api/input-assist/rpc;
  input.right     补 / 校 toggles                     ≤0.1.1: rpc.handle /input-assist)───▶
  fetch POST /api/input-assist/stream                complete   → FIM /beta/completions (non-stream)
    ↳ SSE frames drive ghost progressive render      proofread → LLM layer (llmOnly)
  settings.plugin.item  settings card (Plugins → Plugin Settings)   cancel → abort in-flight requests (RPC & stream alike)
  useInput reads draft · inputActions.setDraft writes back          stream route → FIM stream:true, frames {delta}/{done}
  Mirror layer (mounted on body): red errors in text + gray ghost suggestion after the cursor (rAF-batched render)
  Dictionary layer runs locally (single source src/proofread-dict.ts, bundled into the client at build time)
  Ships its own snapshot store (src/snapshot-store.ts, satisfying the host ObservableSnapshot contract)

Status

• Research & feasibility (2026-08-23)
• First release: dual-side plugin scaffold + completion suggestion bar + typo checking (dictionary + LLM layers, click to fix)
• Typo-check redesign (in-text red underline, per-item navigation, mark-as-correct, shortcuts)
• Inline completion redesign (ghost text, Tab word-by-word accept, dictionary layer moved into the browser at 200 ms, LLM/completion unified 800 ms)
• TypeScript rewrite (all of src/ in TS, tsdown dual-entry build, single-source dictionary)
• CI/CD & release (dual-name npm packages, Trusted Publishing with no token, GitHub Pages website)
• Settings card (Settings → Plugins → Plugin Settings: 11 staged-editable options, save/discard, local validation)
• Model catalog fetching (models.list RPC + dropdown suggestion combobox; default model migrated to deepseek-flash)
• Externalized dictionaries & custom dictionary (data/ plain-text lists + gen-dict generator + CI sync guard; browser localStorage custom dictionary edited in the settings card)
• English spell checking (misspelling + proper-noun dictionaries, word-boundary scanning; LLM layer proofs mixed Chinese/English)
• Real-Chrome UX acceptance: shortcuts field-tested (passed 2026-09-13; clicks, settings card, dictionary section verified on a real dsh on 2026-09-12)
• In-flight request cancellation (retyping re-schedules / Esc / toggling off aborts pending completion and LLM requests; late responses never resurrect and never burn tokens; verified on real hardware 2026-09-13, including Esc inside the debounce window)
• Custom-dictionary import/export (same .txt format: import merges into the staged list, export downloads / copies to clipboard; verified 2026-09-13)
• Streaming progressive render (SSE first-token progressive display + mid-stream cut-off; code & tests done, accepted on real hardware 2026-09-20: SSE frames driving ghost growth, mid-stream Tab accept)
• dsh 0.1.5-rc.2 compatibility (2026-09-15: cordis webServer injection, client-runtime-less snapshot store takeover, RPC over fetch dual path; compatible with the old 0.1.1 runtime — 116 tests green + real-machine curl smoke)
• Composer contenteditable compatibility (2026-09-20: the 0.1.5 composer switching to a Lexical contenteditable silently broke ghost text; findComposer for both runtimes, cursor block-level walk, mirror layer / Tab accept / keyboard interception adapted end-to-end, accepted in a real browser — see docs/17)

Release process

Pushing a v* tag triggers automatically (version must match package.json):

• Release npm — after tests pass, one publish under both names: dsh-input-assist (bare) + @honlnk/dsh-input-assist (scoped mirror), identical content. Auth via npm Trusted Publishing (OIDC), no NPM_TOKEN; already-published names are skipped (idempotent).
• Deploy Pages — site files under docs/ deploy to GitHub Pages.

One-time setup: a Trusted Publisher entry on npmjs.com for each of the two package names (repo honlnk/dsh-input-assist, workflow release-npm.yml, environment empty); GitHub Settings → Pages → Source set to “GitHub Actions”.

git tag v0.1.0 && git push origin v0.1.0   # that's the release

Community

• Show Your Plugins! thread — bilingual intro post in the dsh upstream discussion board, feedback welcome
• Blog: Copilot-style input for DeepSeek Harness — a user-perspective feature tour (Chinese)
• More dsh plugins: GitHub topic dsh-plugin

Uninstall

dsh plugin --profile web remove dsh-input-assist