dsh-composer-predict

Next-message predictions for DeepSeek Harness — one quiet suggestion at the end of every turn, accepted with Tab.

It is the DSH counterpart of Codex's Composer predictions: after a turn completes, the model reads the conversation and guesses what you are about to say next. The guess is rendered as a single faint line in the transcript tail, right under the reply it followed, and several of them can coexist down the history.

What it does

• One line per turn, in the flow — mounted on conversation.chat.turnTail, so it inherits the message column's width and scrolls with the transcript. No banner, no border, no background.
• Tab to accept — the text becomes an editable draft; it is never sent for you. On a non-empty draft it is inserted at the caret instead of replacing what you already typed.
• Esc to dismiss a single suggestion.
• Nothing when there is nothing — with no prediction the entry renders null: no wrapper, no reserved height. A turn in progress counts as nothing, so the previous turn's guess never sits beside fresh output.
• Quiet by default — 30-minute expiry, and only the newest 10 turns keep a line.

Cache behaviour

This is the part worth reading before you tune it.

1. The auxiliary call is a genuine prefix of the conversation. It replays the session's own system prompt, tools and message history, then appends its directive as the final user message. That is the same construction compaction-basic uses, and it is what makes the provider's warm KV prefix cache reusable instead of invalidated.
2. It never invents a route. Provider and model come from session.requestHeader().config. A different route shares no prefix cache at all.
3. It never invents a system prompt, a purpose, or sampling overrides. Rewriting message 0 drops the prefix cache immediately.
4. Results are cached per turn. Key = <log seq>|<provider>/<model>, stored under sessionId → turn. Re-entering a finished turn, or a duplicate turn/end, costs zero model calls. The host keeps 20 turns per session and 64 sessions.

The listener subscribes to session/event with { global: true }. That is required rather than defensive: a plugin fiber is not on a session's scope ancestor chain, so without it the listener never fires.

Install

dsh plugin install link:/absolute/path/to/dsh-composer-predict

Or add it to your profile's package.json:

{
  "dependencies": { "dsh-composer-predict": "link:../../plugins/dsh-composer-predict" },
  "dsh": { "profile": { "bundles": ["dsh-composer-predict"] } }
}

The host half needs a host restart to mount (it serves /plugins/dsh-composer-predict/state); the browser half reloads over HMR.

Configuration

The bundle patch accepts two optional overrides:

- insert:
    - id: composer-predict
      name: dsh-composer-predict
      config:
        maxTokens: 200
        timeoutMs: 25000

Self-check

curl -s 'http://127.0.0.1:<port>/plugins/dsh-composer-predict/diag'
# {"mountedAt":...,"turnEnds":12,"skipped":2,"starts":10,"failures":0,
#  "lastFailure":"","lastKey":"<seq>|<provider>/<model>","sessions":1,"cached":10,"inflight":0}

turnEnds counts completed turns seen, starts predictions actually run, failures auxiliary calls that failed, cached stored predictions. It never carries conversation text.

Requirements

• DeepSeek Harness >= 0.2.0
• A configured provider/model route on the session (the prediction reuses it)

License

MIT