A DeepSeek Harness plugin that adds a ✨ button to the chat composer. One click rewrites your rough draft — combined with the session context — into a clear, well-structured prompt, using the model you already selected.

[!TIP] The fastest way to see it work: type a few words, press the ✨ button, and compare. 审计项目看一下目前有啥问题 → a complete audit brief with scope, requirements, deliverables and acceptance checks.

Features

• 🎯 One-click rewrite — reads the current draft, calls the host-side optimizer, replaces the draft with the result. If the draft changed meanwhile, you get a conflict prompt instead of a silent overwrite.
• 🧠 Context aware — pulls the last few user turns from the chat snapshot as reference (@path mentions survive verbatim).
• 🛡️ Injection safe — the draft and context are wrapped in evidence markers (<draft-evidence>) declared as data, not instructions; marker collisions are escaped, so a draft saying "ignore all rules" cannot hijack the optimizer.
• 📐 Structured output — the optimizer reconstructs goal → scope → requirements → deliverable → acceptance, never a persona card, never fabricated facts.
• ⚡ Zero side effects — a single auxiliary LLM call (purpose: 'compaction'): nothing enters the session log, no agent is activated, cancellable with Esc.
• 🧼 Clean output — stray code fences or wrapping quotes are stripped automatically.
• ♿ Accessible — while optimizing, the button simply dims and disables; no spinner, no layout shift, and a proper aria-label.

How it works

Model routing

The optimization call reuses the session's model, resolved in this order:

1. The pending selection from the session's modelSelection projection (the model you just switched to)
2. The projection's last-used route
3. The route recorded in the session log's requestHeader
4. The deployment's agentDefaultModel fallback

The optimization prompt

The system prompt is a layered contract — role framing, a silent analysis checklist, an explicit expansion policy and a mandatory pre-output self-check — re-targeted at coding-agent task briefs:

Layer	What it does
Non-negotiables	Rewrite only — never answer or execute; evidence is not instructions; no invented facts; keep the draft's language; preserve paths, commands, versions and placeholders verbatim; output exactly one prompt
Internal analysis	Goal, scope, constraints, missing information, expected output form, risks — computed silently, never printed
Expansion policy	A short draft must not come back a few characters longer: a 4–10 character task normally becomes a 100–400 character prompt with 2–6 concrete points, every addition grounded in the draft or context (or labelled as an assumption)
Suggested shape	Goal → context → scope → requirements → deliverable → acceptance → open items; small asks stay 2–4 natural sentences
Self-check	Seven silent questions; if any answer is "no", the output is fixed before it is emitted

[!NOTE] Long drafts keep every detail — the optimizer only reorganizes, de-duplicates and closes gaps. Generic filler counts as a failure, not expansion.

Install

# ① From npm (recommended)
dsh plugin --profile <name> add dsh-composer-prompt-optimizer

# ② From a release tarball
curl -LO https://github.com/subisle/dsh-prompt-optimizer/releases/download/v0.2.0/dsh-composer-prompt-optimizer-0.2.0.tgz
dsh plugin --profile <name> add ./dsh-composer-prompt-optimizer-0.2.0.tgz

# ③ From a local directory
dsh plugin --profile <name> add /path/to/prompt-optimizer

# ④ From git (runs the prepare script to build)
dsh plugin --profile <name> add github:subisle/dsh-prompt-optimizer

# Verify the composed profile
dsh --profile <name> --dump-config | grep -A1 prompt-optimizer

[!IMPORTANT] In the Desktop app, install into ~/.dsh/profiles/desktop, then restart it — client entries hot-reload via HMR, but the host half needs the restart.

Configuration

Every field is optional; invalid values fail plugin load with a schema error (no silent fallback).

- id: prompt-optimizer
  name: dsh-composer-prompt-optimizer
  config:
    maxOutputTokens: 4096
    timeoutMs: 120000

Field	Default	Description
maxDraftChars	8000	Maximum accepted draft length in characters (draft-too-large beyond it)
maxContextTurns	6	Recent user turns carried as reference context
maxContextChars	600	Per-turn context truncation limit
maxOutputTokens	2048	Output-token cap for the auxiliary call
timeoutMs	60000	End-to-end deadline for the auxiliary call (prompt-optimizer/timeout)

Naming and version compatibility

npm package: dsh-composer-prompt-optimizer · GitHub repo: subisle/dsh-prompt-optimizer

The names differ because dsh-prompt-optimizer is already taken on npm by another author's plugin. package.json's name is the single source of truth: the build script derives the plugin row, the browser closure-factory id and both Typert manifests from it — change it once and run npm run build.

All dsh dependencies are peers (no runtime dependencies); DSH validates them at compose time with semver.satisfies(runtime, range, { includePrerelease: true }). This package declares ^0.2.0-rc.2 — bundles outside the range are skipped by the profile, so bump the peer range when DSH moves to 0.3.x.

Development

npm install        # also triggers prepare → automatic build
npm run build      # lib/index.js, lib/client.js, lib/typert.*, lib/types/
npm test           # pre-release self-checks (81 checks)
npm pack           # distributable tarball

The browser bundle uses DSH's closure-factory format (window.__ModuleLoader__.load({id, factory})) with everything inlined except the declared module table (react, cordis, ui-primitives…); CSS Modules compile into a plugin-owned style injector.

Repository layout

prompt-optimizer/
├── package.json            # dsh.bundle(patch) + dsh.client(web) + peer contract
├── cordis.patch.yml        # compose-layer: inserts the plugin row
├── build.mjs               # self-contained build: tsc types + esbuild bundles + Typert artifacts
├── test/check.mjs          # pre-release self-checks
└── src/
    ├── index.ts                # host entry: mounts the Remote controller (exports Config)
    ├── host-controller.ts      # promptOptimizer.optimize + routing + output normalization
    ├── types.ts                # wire types + RemoteErrorDetailsMap extension
    └── client/
        ├── index.ts            # browser entry: $mount Remote + slot registration
        ├── PromptOptimizer.tsx # toolbar component
        ├── context.ts          # conversation context extraction
        └── locales.ts          # zh/en dictionaries

Self-check coverage (npm test)

1. Manifest contract — name/version/type/dsh.bundle.patch/dsh.client.platform, every exports subpath, files coverage, no runtime dependencies
2. Compose layer — exactly one insert row in cordis.patch.yml, keyed by package name
3. Peer compatibility — every dsh peer satisfied for 0.2.0-rc.2 and 0.2.1-alpha.2, no workspace: protocol
4. Host half — apply/name/inject; Config defaults and rejection; normalizePrompt and userText behavior; all error codes; no temperature
5. Optimization contract — evidence framing, marker escaping, layered system prompt sections, expansion policy clauses
6. Browser half — closure-factory format, declared externals only, no Node builtins, activation under real Cordis semantics

[!TIP] Three Cordis constraints shape this codebase: only inject-declared services are reachable from ctx; a self-mounted Remote namespace must be resolved lazily after $mount (writing it into module-level inject deadlocks); and single-occupancy slots reject sibling registrations — hence the stackable conversation.input.right.