Prompt Forecast
English | 中文
Ghost-text prediction of your next action in the DeepSeek Harness Web composer,
in the style of Codex and Claude Code prompt suggestions.
After a completed turn the Host half asks the session's own model for the single
most likely next actionable message, publishes it through a client-visible
session projection, and the Client half renders it as gray text on the first line
of an empty composer. Press Tab or → to accept it into the draft, or
Esc to ignore that suggestion. Typing anything hides it; the text you typed is
never replaced.
The default mode is a Codex-style context fork: the parent's current text
transcript plus one synthetic user turn, Suggest my next message., with
explicit rules that restrict the answer to an action.
Requirements
- DSH
0.2.1-rc.2 or newer within 0.2.x, declared as engines.dsh: ^0.2.1-rc.2.
- This plugin is built on the 0.2.1-rc.2 client and Host surfaces: the
conversation.input.overlay composer slot, InputActions.captureInsertion() /
insertText(), and the session-projection carrier fed by agent/turn-stopping.
dsh plugin add validates the range during installation and restores the
previous profile state when it does not match; profile startup checks it
again. An unmet range refuses to load the plugin; upgrade DSH or install a
plugin release whose engines.dsh matches your runtime. A DSH 0.3 release
will need a matching plugin release rather than an exemption.
What it does
- Fork mode (default) predicts from the full text transcript; digest mode uses a
compact JSON excerpt.
- Predicts the next action, not a status report or an outcome guess.
- Writes in the language the user writes in, one short line.
- Uses the route that served the turn (falling back to the profile default
model), so no extra provider setup is required.
- Skips subagent sessions and turns with no human input.
- Fails silently: a slow, aborted, or failed prediction never surfaces an error
in the conversation.
What gets predicted
The target is the next actionable message, not a literal continuation of the
conversation. A useful suggestion is one the user would send to tell the
assistant what to do next:
- an instruction: run the tests, fix this error, check the projection cache,
explain the failure
- a choice or approval: use option A, go ahead, keep it
- a correction: no, change the client side
- a request to explain or diagnose: why does the fork call time out?
- a short report followed by an instruction: restarted, still empty — check the
logs
Pure status reports, test results, log observations, narration, and guesses about
what might happen are rejected; NONE is a valid answer. A conservative
post-filter drops obvious pure-report clauses, and fork mode falls back to the
digest when the result is unusable.
Prediction modes
mode: 'fork' (default) mirrors the Codex composer-prediction shape. It takes
the session's current text transcript — system and developer instructions plus
user/assistant text — drops tool calls, tool results, reasoning, and attachment
blocks, then appends one synthetic user turn:
Suggest my next message.
The model's reply becomes the ghost line. No child session or session event is
created; the fork is an in-memory message list for one auxiliary call. A short
action-rules block is appended to the inherited system prompt so the model writes
the user's next instruction rather than a status report. Because the whole
transcript is sent, this mode costs more tokens and needs a larger timeoutMs
than the digest mode.
mode: 'digest' keeps the earlier lightweight path: a compact JSON excerpt
(recent_user_messages, latest_assistant_reply, tools_used) with a short,
action-oriented system prompt. Use it for cheap predictions or long sessions
where sending the transcript is too expensive.
Configuration
The row's config block in cordis.patch.yml overrides every default:
| Field | Default | Meaning |
|---|
enabled | true | false leaves the Host half unregistered |
mode | fork | fork = full transcript + Suggest my next message.; digest = compact JSON excerpt |
maxTokens | 1024 | Output budget of the prediction request |
temperature | 0.4 | Sampling temperature |
timeoutMs | 12000 | Hard bound on the request; a slower answer is dropped |
maxPromptChars | 80 | Clip length of the predicted prompt |
maxForkChars | 12000 | Character budget of the forked conversation window (system/developer text is always kept) |
fallback | true | When the fork yields no text, retry with the compact digest |
maxInputChars | 6000 | Digest-mode excerpt size bound |
skipSubagents | true | Subagent sessions get no prediction |
provider | turn route | Explicit provider override |
model | turn route | Explicit model override |
The prediction is produced while the turn boundary is still open
(agent/turn-stopping), so its text arrives with the finished turn instead of
requiring a poll. timeoutMs is the price: a prediction that does not settle
within that budget is dropped and the turn finishes without one.
Install
From GitHub (no build step)
dsh plugin --profile web add github:littepointR/dsh-prompt-forecast
dsh web
dsh plugin add forwards to pnpm in the chosen profile, installs the package,
and automatically adds the bundle to dsh.profile.bundles because the package
declares dsh.bundle.patch. Restart the profile for the plugin to load. The
package is plain JavaScript with no prepare build script, so a git install
needs no extra pnpm build approval.
To pin the tagged release:
dsh plugin --profile web add github:littepointR/dsh-prompt-forecast#v0.2.4
The Desktop application's desktop profile is reserved for the Electron app, so
the CLI rejects --profile desktop. Use Settings → Plugins there, or install
into a CLI profile.
From a local checkout
Use Plugin Manager with this package directory as the bundle target. The bundle
patch inserts the prompt-forecast row; it survives restarts because it is a
normal profile bundle.
Limits
- Only an empty composer shows a suggestion. Completing a partially typed draft
needs caret geometry the slot does not expose yet and is not part of this
version.
- The suggestion is a hint, not an action: accepting inserts plain text into the
draft and nothing is sent.
- Prediction quality depends on the session model; a very small model may return
a generic follow-up.
Development
node --test test/host.test.mjs test/client.test.mjs
node --check index.js && node --check client.js
index.js — Host half: generate at agent/turn-stopping, publish through the
promptForecast projection.client.js — Client half: register in conversation.input.overlay, render the
ghost line, handle Tab/→/Esc.test/host.test.mjs — text normalization, digest building, projection fold.test/client.test.mjs — structural smoke test of the Client render branches
(no DOM emulation).
License
MIT
