dsh-composer-suggest

Composer suggestions for the DeepSeek Harness Web UI.

• Next message: after the agent replies, the empty composer shows the message you are most likely to send next.
• Draft continuation: with text in the composer and the caret at its end, press Tab. The composer shows "Continuing…" with a spinner while the model works from the conversation above and your draft; press Tab again to accept the result, or Esc to cancel. Nothing is asked until you press Tab, and Tab keeps its usual job anywhere else (empty composer, caret mid-draft, a / command or @ reference).
• Tab accepts a suggestion as an editable draft (nothing is sent), Alt+Down / Alt+Up cycle up to three candidates, Esc withdraws the offer, and typing discards it.
• Settings → General has one switch per mode, and a status line that shows which model route answered last, how long it took, or why a request failed.

Every suggestion is one model call and uses tokens. It is not written to the Session log and does not show in the token meter.

Install

From a checkout:

git clone https://github.com/WuweiRE/dsh-composer-suggest.git
dsh plugin --profile <name> add ./dsh-composer-suggest

From npm, once a release is published there:

dsh plugin --profile <name> add dsh-composer-suggest

Restart the Harness afterwards, then refresh the Web page. Developed and tested against DeepSeek Harness 0.2.2-rc.1. It needs a Harness that provides conversation.input.overlay, settings.general.item, connection.admit(), ctx.llm and session.deriveMessages().

Model

By default a suggestion is answered by the model the Session already uses, at that model's lowest reasoning effort, with no temperature set.

To answer with a specific model instead, override the row in your profile's cordis.patch.yml. Any option you leave out keeps its default:

- id: composer-suggest
  name: dsh-composer-suggest
  config:
    model: claude-sonnet-5-5

Without a provider, the model is looked up among the providers mounted in your profile, trying the Session's provider first. If none offers it, the request fails with model-unavailable instead of silently using another model. Set provider and model together to name the route exactly. See cordis.patch.yml for every option.

Credits and licenses

This plugin is a combination of ideas and code from two MIT-licensed DeepSeek Harness plugins. Their license texts are kept in this repository, and THIRD_PARTY_NOTICES.md lists what comes from where.

Project	Author	License	Used here
dsh-composer-predictions	Stijnus	MIT, LICENSE.stijnus	Code adapted: the streaming Host endpoint and owner check, prompt construction, candidate normalization, caret measurement, ghost positioning, SSE parsing and the keyboard handling.
dsh-composer-prediction	Amir Bazylkhanov	MIT, LICENSE.bazylkhanov	Behaviour reimplemented from its documentation and source: withdrawing a dismissed offer for the rest of the turn, IME handling, reporting a failed request in Settings, and using the model's cheapest advertised reasoning effort.

What this project adds: continuation is triggered by Tab instead of running on every typing pause, a visible "Continuing…" state with cancel, separate switches for the two modes, model selection that follows the Session or pins a model, a status line in Settings, and short on-screen notices for empty or failed continuations.

This project is released under the MIT License, Copyright (c) 2026 WuweiRE, see LICENSE. It is not affiliated with DeepSeek or with the authors above.

Test

node --test