DSH Plugins Marketplace

DSH Plugins

Plugins

/

dsh-next-step-suggest

A

dsh-next-step-suggest

Manifest valid

Ghost-text next-message suggestion for the DeepSeek Harness composer. One suggestion, two keys: Tab accepts, Esc dismisses, nothing is ever auto-sent.

UI (client)hasBundlePatch

Next-Step Suggest

A DeepSeek Harness plugin that shows your next message as ghost text inside the composer. Press Tab to accept — it fills the draft and never sends.

One suggestion, two keys, three minutes apart. Deliberately quiet.

Why it exists

Most "next prompt" plugins guess what you would type. That produces continue, ok, next — noise.

This one asks a different question: what should the task advance to next? When the answer is already in your session state, it does not call a model at all.

How it decides

rule hit?  ── yes ──►  "Continue: <your in-progress todo>"
    │                  "Next: <your pending todo>"
    │                  "Continue with the plan"
    no
    ▼
model call (last 8 turns + todos + cwd)  ──►  one short instruction
  • Rules first: in-progress todo → pending todo → active plan.
  • Model fallback: asks for the next concrete action, in your language, never a bare "continue".
  • Vacuous guard: a suggestion with no object (continue, fine, ok) is dropped server-side.

Install

Until it is on npm, install from a local checkout:

dsh plugin --profile <your-profile> add /path/to/next-step-suggest

Or copy the directory into your profile's node_modules and add the row by hand — see cordis.patch.yml.

Requirements: DSH 0.2.1-rc.x, web UI.

Use

KeyEffect
TabFill the draft with the suggestion (editable, one Ctrl+Z to undo)
EscSilence it for this turn

If you keep typing, the ghost disappears on its own. After accepting, the next suggestion can appear 60 s later.

When it appears

All of these at once:

  1. the session has at least one user + assistant exchange
  2. the agent is idle
  3. the composer is empty (whitespace counts as empty)
  4. no suggestion is already on screen
  5. frequency gates pass — 60 s retry gap, 3 min between appearances, 6 appearances per session, 2 consecutive Esc mutes the session

It shows 1.2 s after the reply arrives, so you can read first.

Nothing showing? The console prints [next-step] skip: <reason> once per reason change — running, draft, hidden, showing, retry-gap, cooldown, session-cap, or muted names the gate that blocked it.

Configuration

Per-row config in your profile's cordis.patch.yml:

- id: next-step-suggest
  name: 'dsh-next-step-suggest'
  config:
    enabled: true        # false = silent, no uninstall needed
    rules: true          # false = always ask the model
    candidates: 1        # 1–3
    maxTokens: 120
    temperature: 0.9
    provider: ''         # empty = follow the session's own model
    model: ''
    reasoningEffort: ''  # empty = unspecified
    dedup: true
    timeoutMs: 6000

Reload the plugin after changing config (remove_bundle then install_bundle).

Design notes

  • Ghost text is a projection, not content. It renders in the composer overlay slot with pointer-events: none; it is never part of the draft, so it cannot be sent by accident.
  • Alignment is measured, not guessed. The ghost copies the native placeholder's rect and font metrics, so it lines up on any theme or density.
  • Keys are arbitrated in the capture phase, and only when the ghost is visible in the same composer card — two panes cannot double-write.
  • Sessions are isolated. Every async result is checked against the originating session id before it touches state; late responses are dropped.
  • Failures are silent. Timeouts, provider errors, empty output: nothing appears, nothing pops up.

Known limitations

  • Two composers open at once: the native placeholder is not hidden (the ghost may overlap it).
  • Server module has ESM caching: after editing code, reinstall the plugin — a page refresh only reloads the client half.
  • Frequency constants live in client.js; tune them by editing the file and reinstalling.

License

MIT — see LICENSE.

Comments

Loading…