Prompt Optimizer (提示词优化)

A DSH plugin that rewrites what you type into the composer into a precise prompt before the model sees it, using one cheap side call on the session's own route.

The rewritten text is what enters the durable log, so the chat bubble and the model always agree: you never see one thing and send another.

Install

dsh plugin --profile <profile> add dsh-prompt-optimizer

Or install the prebuilt tarball from the latest release through the app's plugin manager. No build step, no allowBuilds approval.

Turning it on

The switch is the whale in the composer's tool row (conversation.input.right), between the model selector and Send.

Click it to toggle; the state lives in ~/.dsh/prompt-optimizer/state.json and survives a restart. Hover it to read the last before/after pair for the current conversation.

Two states

state	rendering
off (default)	a dark silhouette used as a CSS mask over currentColor, so it carries no fill colour of its own and follows the host theme
on	the full-colour gold artwork

Both states are inlined into client.js as data URIs. The chip is a 16px control that must paint on the first frame, and a round trip per state would flash an empty button. assets/ keeps the source PNGs the constants are generated from.

What it does

While it is on, every message you send that looks like prose gets one non-streaming call that turns it into a structured prompt: goal first, then the constraints and acceptance criteria already implied by what you typed. Concrete details — paths, names, numbers, error text — are preserved, chit-chat is dropped, and the original language is kept.

What it never touches

• Slash commands (/compact, …), messages shorter than minChars, messages longer than maxInputChars, and image-only messages.
• A message it already rewrote inside the same turn (identified by message id).
• Anything when the side call fails: a provider error, a timeout, or an empty rewrite sends the text exactly as you typed it, and the failure is logged as a warning (dsh-prompt-optimizer: rewrite failed; …).

Where it hooks

agent/pre-step — the one waterfall that decides which messages enter a step. agent/request is deliberately not used (the loop refuses message changes there), and a request-side hook would rewrite a message already in the durable log, so the transcript would disagree with the model.

The listener spreads the incoming decision ({ ...decision, messages }), so fields such as startsRequestSeries and notices added by later listeners survive.

Configuration

cordis.patch.yml in this bundle is the only config surface:

- insert:
    - id: prompt-optimizer
      name: dsh-prompt-optimizer
      config:
        enabled: false

The switch in the composer writes the state file, which wins over this row; enabled here is just the initial posture. The rest of the knobs live in the state file and can be edited by hand:

field	default	meaning
enabled	false	the switch
provider / model	""	pin the rewrite route; empty = the session's own route
minChars	8	shorter messages are sent untouched
maxInputChars	6000	longer messages are sent untouched
maxOutputTokens	1200	cap for the rewrite call
timeoutMs	12000	deadline for the rewrite call
history	20	how many rewrites the hover panel can read back

HTTP surface

GET /prompt-optimizer/state[?sessionId=…] → { ok, enabled, minChars, route, last }

POST /prompt-optimizer/state with { "enabled": true } → { ok, enabled }

Both are served by the client half's toggle. There is no remote RPC: the route is tiny and does not need one.

Display

• Title / description: locale/en.json, locale/zh.json (meta.title, meta.description).
• Icon: icon.png (256×256, the gold whale on a dark tile), exported as ./icon.

License

MIT