dsh-composer-autocomplete
Manifest validTab autocomplete for the DeepSeek Harness composer: inline ghost text, next-message suggestions and early candidate acceptance. MIT community enhancement.
DeepSeek Harness Composer Autocomplete
English | 简体中文
dsh-composer-autocomplete adds inline ghost-text completion and next-message suggestions to the native DeepSeek Harness Web composer. Tab accepts, Esc dismisses, and Alt+↓ / Alt+↑ cycles candidates. Acceptance inserts an editable draft; it never sends a message.
Licensed under MIT. By default, predictions use the model configured in your session and consume that service's tokens.

The screenshot shows the actual Harness UI with a deterministic local test model.
Quick install for Web · Community walkthrough (中文)
If this makes writing in Harness easier, a GitHub Star helps others find it.
Install or update
Requirements: Node.js ≥ 24.9 and DeepSeek Harness. Tested with Harness 0.2.1-rc.2 Web. Install the prebuilt package from GitHub Releases:
dsh plugin --profile web add https://github.com/liangzhiyu/dsh-composer-autocomplete/releases/download/v0.2.3/dsh-composer-autocomplete-0.2.3.tgz
dsh web
Restart any running Harness instance after installation or updating. For the official desktop app, use its plugin installation UI to select the local package, or use the bundled dsh command to manage the desktop profile. The desktop shell has not been tested; this plugin uses the native composer slot on the Web client.
Usage
- Empty composer: suggests your next message after an assistant turn finishes.
- Draft completion: requests a continuation after a typing pause, with the caret at the end. The default pause is 150 ms.
- Fast mode (default): one candidate, less context and a shorter output budget. Full mode keeps multiple candidates and the configured context limits.
- Tab: accepts the first complete candidate immediately, while later candidates may still be generating. Send the message yourself when ready.
- Alt+↓ / Alt+↑: cycles through up to three candidates in Full mode.
- Esc: dismisses the current suggestion. Editing the draft allows a new prediction.
- Settings → General → Composer predictions: controls the master switch, prediction modes, Fast/Full, and the typing pause (100–5000 ms).
Preferences are stored in the current browser on this device. IME composition keeps ownership of Tab. Predictions pause while the agent is running and stay out of native /command and @reference menus. Draft edits, session changes, disabling predictions, and hiding the page cancel pending requests. Text insertion goes through Harness's public input actions; a refused insertion never replaces the draft.
The native caret remains visible while suggestions are shown. After acceptance, it stays at the end of the inserted text; click or use the arrow keys to keep editing.
Switching sessions clears suggestions before paint and resets IME and dismissal state. Every candidate belongs to one session, draft, and request, so late responses cannot appear in or be accepted by another view. Result caches are isolated by session ID, session sequence, prediction mode, draft, and Fast/Full setting. Returning to a session can reuse its own valid cache; a new conversation turn invalidates the previous revision.
Typing the exact prefix of an already completed candidate reuses its remaining text locally, with no new model call. Unrelated edits invalidate it. IME composition pauses suggestions and keeps ownership of Tab.
Existing saved typing pauses are preserved. The 150 ms default applies only when no value has been saved. After upgrading, select 150 ms in settings if an older 300 ms preference remains.
Latency
Measurements use the same deterministic local test model, timed from the end of typing. Each row is the median of six samples:
| Version / setting | First ghost text | Ready for Tab |
|---|---|---|
| 0.2.2, 300 ms pause | 405 ms | 426 ms |
| 0.2.3 Fast, 150 ms pause | 256 ms | 277 ms |
| 0.2.3 Fast, 100 ms pause | 205 ms | 235 ms |
Matching an existing candidate measured 11 ms to update, with no network request (median of six typed characters).
These measure the local UI/SSE pipeline, not DeepSeek cloud speed or output quality. Actual latency depends on the network, model time to first token, formatting, and generation speed. See method and raw data. A shorter pause may increase request frequency; choose a longer pause in settings if preferred.
Model usage and configuration
Predictions reuse the session's configured model route and host authentication; the plugin stores no separate API key. Recent conversation, tool activity, and the current draft are sent to that model and consume tokens. These auxiliary calls do not appear as conversation messages or in the session's displayed token usage. Configure an available model in Harness before using predictions.
Override host parameters in your profile's cordis.patch.yml:
For a slow thinking model, a separately configured fast prediction model can help: set both provider and model below. This leaves the main conversation model unchanged. Fast mode caps the configured limits at one candidate, 6 messages, 4000 context characters, 2 tool notes and 96 output tokens; Full mode uses the configured limits.
- id: composer-autocomplete
config:
maxOutputTokens: 220
timeoutMs: 20000
maxMessages: 10
maxChars: 8000
maxSuggestions: 3
maxToolNotes: 6
cacheTtlMs: 60000
maxCacheEntries: 200
# Set both to use a dedicated prediction model; otherwise use the session route.
# provider: your-provider
# model: your-model
# reasoningEffort: off
Development and verification
The plugin uses native ESM and a Harness browser module. It has no build step or production npm dependencies.
node --test test/*.test.mjs
node tools/check-package.mjs
pnpm pack --pack-destination artifacts
Browser checks use Playwright against an isolated Harness Web instance with the plugin and local test-model adapter installed:
pnpm install --frozen-lockfile
DSH_URL=http://127.0.0.1:3087 node tools/browser-smoke.mjs
DSH_URL=http://127.0.0.1:3087 node tools/session-switch-smoke.mjs
DSH_URL=http://127.0.0.1:3087 node tools/caret-smoke.mjs
DSH_URL=http://127.0.0.1:3087 node tools/speed-smoke.mjs
The caret check captures the native caret without hiding it. With Pillow installed, verify its actual screenshot pixels with python3 tools/verify-caret-pixels.py.
See verification scope and setup and changes.
Uninstall
dsh plugin --profile web remove dsh-composer-autocomplete
Restart the relevant instance. Browser preferences remain; remove the dsh-composer-autocomplete.* local-storage keys in developer tools to reset them.
Comments
Loading…
Similar plugins
by WuweiRE
Composer suggestions for DeepSeek Harness: next-message ghost text after a reply and Tab-triggered draft continuation.
★ 0
MIT
JavaScript
Nov 18, 2026
dsh plugin --profile web add dsh-composer-suggestby Andy-zch
Ghost-text next-message suggestion for the DeepSeek Harness composer. One suggestion, two keys: Tab accepts, Esc dismisses, nothing is ever auto-sent.
★ 0
MIT
JavaScript
Nov 20, 2026
dsh plugin --profile web add dsh-next-step-suggestby strangeyu911-tech
Inline ghost-text next-word prediction for the DeepSeek Harness composer
★ 0
MIT
JavaScript
Nov 12, 2026
dsh plugin --profile web add dsh-composer-autocompleteby littepointR
Ghost-text next-prompt prediction in the DeepSeek Harness Web composer, accepted with Tab.
★ 0
MIT
JavaScript
Oct 27, 2026
dsh plugin --profile web add @local/dsh-prompt-forecastby QingJ01
Next-prompt ghost text predictions for DeepSeek Harness; Tab or ArrowRight fills an editable draft.
★ 4
↓ 160/wk
MIT
JavaScript
Nov 5, 2026
dsh plugin --profile web add dsh-prompt-predictionby sakura-hly
Codex-style composer for the DeepSeek Harness Web GUI: prompt history, Tab completion, and context-aware next-prompt suggestions.
★ 0
MIT
JavaScript
Nov 4, 2026
dsh plugin --profile web add dsh-codex-composer