dsh-prompt-seed
DSH composer：主动进行一次调用的草稿改写和独立的下一条回复预测，支持会话范围内的任务和安全的草稿恢复。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Zian-anson/dsh-prompt-seed#69ad427f935653f0c5ae0acb27227d02cdc6f415Prompt Seed 以 dsh-prompt-seed 发布，当前版本为 0.20.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
One contextual DSH composer entry: ✦ Optimize when the draft has content; 🔮 Crystal ball when empty. The tools improve the draft or predict the user's most likely next reply. Neither sends messages.
Version 0.20.0. See acceptance evidence for model quality, real-UI verification and known limitations.
Short drafts gain essential execution details, boundaries and proportionate verification. Long spoken drafts become organized and shorter. Precise requests stay concise. Preserve intent, scope, facts, authorization, language, numbers, technical names and negative requirements—including “discuss only, do not implement.” Do not invent features, metrics or an unchosen stack.
After generation, Another version uses the round's original draft, not the previous rewrite. Manual edits start a new round on the next click. Use ⋯ → Restore original directly while untouched; after editing, preview it before explicitly replacing the draft. Empty input, greetings and pure confirmations make no call. Unchanged results say “Unchanged,” distinct from improvement and failure.
Crystal ball is available when the composer is empty, the conversation has an assistant reply and the assistant is idle. It uses recent conversation and user style to suggest one short reply: acknowledgement, question, correction, extra information or continuation. With insufficient evidence it reports “Unable to predict for now.” Blank conversations make no call. After a prediction fills the draft, the main icon switches to Optimize. ⋯ offers Another reply and Withdraw while the candidate is untouched; editing removes those actions.
Success has no persistent text; a spinner indicates progress. Failure and unchanged results use short labels with details on hover. ⋯ appears only when recovery, background cancellation or a retained result is available. Switching the icon does not cancel either feature’s job.
Each feature owns its prompt, context strategy, controller, task state, result and undo. Only connection, task storage, timeout and safe-write primitives are shared.
Published release:
dsh plugin --profile <name> add \
https://github.com/Zian-anson/dsh-prompt-seed/releases/download/v0.20.0/dsh-prompt-seed-0.20.0.tgz
For a local candidate, run npm pack, then dsh plugin --profile <name> add ./dsh-prompt-seed-0.20.0.tgz.
Keep the previous package for recovery and restart DSH while idle after Host changes. Use the compatible CLI bundled with Desktop. Do not edit the installed copy.
Remove with dsh plugin --profile <name> remove dsh-prompt-seed.
config:
route: /api/prompt-seed/optimize
# provider / model: specify both, or follow the Host default selection
requestTimeoutMs: 20000
context: true
samples: true
templates: true
User templates in $DSH_HOME/prompt-seed/prompts/ remain intact: optimize.md takes precedence, then legacy seed.md.
Existing clarify, precise and chat files are preserved but are not selected by the single rewrite contract.
Local sample logs contain input/output previews and are never automatically uploaded; samples:false disables them.
POST /api/prompt-seed/optimize remains loopback-only and requires application/json.
{ "action": "optimize", "kind": "optimize", "text": "...", "sessionId": "...", "stream": true }
Actions: optimize / predict / status / bind / cancel / consume. Queries and binding default to optimize when kind is omitted. New clients pass "optimize" or "predict". Cancellation/consumption retain legacy jobId-only calls; an explicit kind validates ownership.
Both generators accept stream:true and return NDJSON meta / delta / final events with jobId and kind.
Nonstream optimize retains its final JSON. Nonstream predict retains {ok:true,text} or {ok:false,code,error} without requiring new fields.
Status adds protocolVersion:2, context availability and the last persisted user-message identifier, not chat content. ?debug=1 adds diagnostics to nonstream optimize.
Failures, empty/truncated results and lost explicit full paths/files/URLs/identifiers never overwrite the draft.
Node ≥22.19; zero runtime dependencies, no install step needed.
npm test # builds lib/, runs 63 tests and documentation/install guards
npm run bench -- --dsh <compatible-dsh-command> --profile web
node tools/quality-judge.mjs bench/latest.json # offline review sheet, no model call
| Validation | Count |
|---|---|
| Real-socket checks | ✅ 13 checks, run by npm test |
src/ is authoritative; rebuild lib/ after editing. Behavioral tests drive the shipped toolbar/controller with controllable drafts, sessions and model streams rather than checking function names.
bench/corpus.json contains 30 rewrite cases; bench/predict-corpus.json contains 10 contextual prediction cases.
The benchmark temporarily loads a development-only plugin into a separate web Host and uses DSH's llm service and templates. It never reads credentials or calls a provider directly. Synthetic context stays in fixtures, not user sessions; the benchmark entry is not packaged.
The test Host must already have a working model connection. Desktop-account evaluation requires explicit temporary-loading approval; see benchmark setup.
Options: --start, --limit, --out, --provider, --model.
Measure both features separately, excluding zero-call cases: p50 ≤3 seconds, p95 ≤8 seconds. At least 80% of improvable rewrite cases should gain useful substance without distortion. Synonyms, politeness and filler do not count; precise originals may remain unchanged. Judge predictions by whether the user might plausibly reply that way, not whether they advance the task. Model evaluation and static checks do not replace real DSH clicks.
A model can be slow or wrong. Idle timeout is 15 seconds; the default total Host budget is 20 seconds. Rewrite context favors newest messages: up to 3 user turns and 2 assistant tails, 1400 characters total. Prediction has its own six-turn window. Without reliable revision/semantic APIs, results remain preview-only.
DSH 0.2.1-rc.2 does not persist unsent image attachments across app restart. Rewriting preserves them, but after restarting, reattach the original images. Text and reference chips use host-owned draft recovery; see acceptance evidence.
MIT.