dsh-composer-predictions 1.2.0 → 1.2.2

package/CHANGELOG.md

@@ -6,6 +6,36 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [1.2.2] - 2026-10-25
+
+### Fixed
+
+- Both prediction routes require Harness connection admission and owner access,
+  including the app's Host/Origin checks, before reading sessions or using tokens.
+- Reject non-JSON, malformed, oversized and invalid-mode requests without model work.
+- Simultaneous requests from different views no longer cancel each other; disabling
+  the host plugin cancels outstanding model calls and clears its cache.
+- Cache draft text exactly, preventing one draft from reusing a hash collision's suggestion.
+- Disconnected or cancelled calls never cache partial output; token-limited output
+  offers only complete lines.
+- Clear a refused insertion's ghost and resume predictions after IME composition ends.
+
+### Changed
+
+- Declare Harness compatibility `>=0.2.1-rc.3 <0.3.0-0` (tested on `0.2.1-rc.3`).
+- Add HTTP admission, input validation, cancellation, cache and IME regressions.
+- The manual probe now requires an authenticated cookie and fails on HTTP errors.
+
+## [1.2.1] - 2026-10-25
+
+### Changed
+
+- Publish desktop/web installation guidance and Chinese usage instructions on npm.
+- Expand package search terms for ghost text and inline autocomplete.
+- Release verification waits for npm's asynchronous publication to become visible.
+- Validate GitHub Actions trusted publishing with provenance. Runtime behavior is
+  unchanged from 1.2.0.
+
 ## [1.2.0] - 2026-10-24
 
 ### Added

package/README.md

@@ -4,7 +4,10 @@
 [![license](https://img.shields.io/npm/l/dsh-composer-predictions)](LICENSE)
 [![topic](https://img.shields.io/badge/topic-dsh--plugin-blue)](https://github.com/topics/dsh-plugin)
 
-AI ghost text for the message box in the DeepSeek Harness Web GUI.
+AI autocomplete and ghost-text suggestions for the composer in **DeepSeek
+Harness desktop and the Web GUI**.
+
+**Current release: 1.2.2** · [Install or update](#install) · [中文说明](#中文说明)
 
 - **Next message** — after the agent replies, the empty composer shows a
   streamed suggestion for your likely next message. **Tab** accepts it as an
@@ -50,7 +53,10 @@ on this device. Other browser origins and devices keep their own preference.
 
 ## Install
 
-Inside the DeepSeek Harness Web GUI: **Plugins → Install**, then enter
+Requires Harness `>=0.2.1-rc.3 <0.3.0-0`; desktop and web are tested on
+`0.2.1-rc.3`. Predictions use the same authenticated owner access as the app.
+
+In the **DeepSeek Harness desktop app or Web GUI**: **Plugins → Install**, then enter
 
 ```
 dsh-composer-predictions
@@ -87,6 +93,20 @@ https://raw.githubusercontent.com/Stijnus/dsh-composer-predictions/main/market-i
 
 ![Inline continuation ghost](https://raw.githubusercontent.com/Stijnus/dsh-composer-predictions/main/docs/continuation-ghost.png)
 
+## 中文说明
+
+**DeepSeek Harness 桌面端 / Web 的 Tab 自动补全与输入预测。** 回复结束后，
+空输入框显示下一条消息的灰字建议；输入停顿时，续写光标末尾的草稿。
+`Tab` 采纳已完成的候选为可编辑草稿，不会自动发送；`Alt+↓/↑` 切换最多三个候选，
+`Esc` 关闭。设置 → 通用 → 输入预测可开关，设置保存在当前设备的浏览器中。
+
+安装或更新：**插件 → 安装 → `dsh-composer-predictions@1.2.2`**，然后重启 Harness。
+CLI 用户可运行 `dsh plugin --profile web add dsh-composer-predictions@1.2.2`，
+把 `web` 换成自己的 profile。预测会调用会话模型并消耗 token，虽然不会出现在会话日志和 token 统计中。
+
+也可以把上方 [Install](#install) 中的市场索引添加到「插件 → 发现」。
+演示、完整配置与限制见本页；[1.2.2 发布说明](https://github.com/Stijnus/dsh-composer-predictions/releases/tag/v1.2.2)。
+
 ## How it works
 
 - The **Host half** registers two routes on the Harness web server:
@@ -107,8 +127,9 @@ https://raw.githubusercontent.com/Stijnus/dsh-composer-predictions/main/market-i
   model restated are trimmed. A continuation keeps the model's leading
   separator space; the Client adds one only when a non-CJK draft and
   continuation would otherwise glue two words together.
-- Results are cached per Session sequence (and per draft for continuations), so
-  a repeated request for unchanged context costs no model call.
+- Completed results are cached per Session sequence (and per draft for
+  continuations), so a repeated request reuses the cached result. Simultaneous
+  requests from different views have independent cancellation and may each call the model.
 - The **Client half** registers into `conversation.input.overlay`, debounces
   requests (600 ms after a turn settles, 500 ms after a typing pause), aborts
   superseded requests, and renders the ghost over the empty composer or at the
@@ -153,8 +174,8 @@ A suggestion longer than 240 characters is cut at a word boundary.
 - A continuation is hidden when there is too little room after the caret; it
   never shifts left over your draft.
 - Final suggestions are announced through an accessible status message.
-- Requires a Harness build that provides `conversation.input.overlay`,
-  `ctx.llm`, and `sessions.deriveMessages()`.
+- Requires a compatible Harness build with `conversation.input.overlay`,
+  `settings.general.item`, `connection.admit()`, `ctx.llm` and `session.deriveMessages()`.
 
 ## Test
 
@@ -162,16 +183,18 @@ A suggestion longer than 240 characters is cut at a word boundary.
 node --test
 ```
 
-Probe a running GUI to watch the stream and cache timing:
+Probe a running GUI to watch the stream and cache timing. First set `DSH_COOKIE`
+to an authenticated Harness cookie in your environment; treat it as a secret
+and never commit or include it in logs.
 
 ```bash
-node tools/live-probe.mjs <sessionId> continue "Can you also"
+DSH_URL='http://127.0.0.1:<port>' node tools/live-probe.mjs <sessionId> continue "Can you also"
 ```
 
-The 30 cases drive both Host endpoints against a fake Session and LLM service:
+The 36 cases drive both Host endpoints against a fake Session and LLM service:
 prompt contents (including tool activity), candidate normalization and spacing,
 streaming events, per-turn and per-draft caching, route and effort selection,
-failure reporting, and request validation. Client helpers (continuation
+failure reporting, HTTP authentication/owner checks, cancellation and request validation. Client helpers (continuation
 joining, SSE parsing, streaming candidates) are unit-tested through the bundle's
 internal `__testing` export. Component regression tests cover streaming acceptance,
 refused insertions, right-edge placement, truncated streams and accessible copy.

package/client.js

@@ -362,6 +362,7 @@ window.__ModuleLoader__.load({
       const running = useSessionStatus((state) => (state === undefined ? false : state.get(sessionId)?.running === true))
       const [ghost, setGhost] = React.useState(NO_GHOST)
       const [box, setBox] = React.useState(null)
+      const [compositionEnd, setCompositionEnd] = React.useState(0)
 
       const abortRef = React.useRef(null)
       const composingRef = React.useRef(false)
@@ -468,6 +469,7 @@ window.__ModuleLoader__.load({
         }
         const onCompositionEnd = () => {
           composingRef.current = false
+          setCompositionEnd(value => value + 1)
         }
         document.addEventListener('compositionstart', onCompositionStart, true)
         document.addEventListener('compositionend', onCompositionEnd, true)
@@ -512,7 +514,7 @@ window.__ModuleLoader__.load({
           void request('next', '')
         }, SETTLE_MS)
         return () => clearTimeout(timer)
-      }, [sessionId, running, draft, request, predictionsEnabled])
+      }, [sessionId, running, draft, request, predictionsEnabled, compositionEnd])
 
       // `continue`: ask after a pause, only with the caret at the draft's end.
       React.useEffect(() => {
@@ -528,7 +530,7 @@ window.__ModuleLoader__.load({
           void request('continue', draft.slice(-MAX_DRAFT_CHARS))
         }, TYPE_PAUSE_MS)
         return () => clearTimeout(timer)
-      }, [sessionId, draft, request, predictionsEnabled])
+      }, [sessionId, draft, request, predictionsEnabled, compositionEnd])
 
       // A prediction skipped while the page was hidden is retried on return.
       React.useEffect(() => {
@@ -682,14 +684,14 @@ window.__ModuleLoader__.load({
         const expected = ghost.mode === 'continue' ? draft + value : value
         dismissedRef.current = { draft: expected }
         abortRef.current?.abort()
+        setGhost(NO_GHOST)
+        setBox(null)
         try {
           // A refused insertion must never be bypassed by replacing the draft.
           if (inputActions.insertText(value, inputActions.captureInsertion()) !== true) return
         } catch {
           return
         }
-        setGhost(NO_GHOST)
-        setBox(null)
         // Accepting leaves an editable draft: hand the keyboard back to the editor.
         try {
           const found = composerElements()

package/index.js

@@ -14,7 +14,7 @@
 
 export const name = 'composer-predictions'
 
-export const inject = ['llm', 'sessions', 'webServer']
+export const inject = ['connection', 'llm', 'sessions', 'webServer']
 
 /** Same-origin endpoint the Client half posts to (JSON, kept for compatibility). */
 const ENDPOINT = '/composer-prediction'
@@ -85,7 +85,13 @@ const CONTINUE_SYSTEM_PROMPT = [
 export function apply(ctx, config) {
   const options = resolveOptions(config)
   const cache = new Map()
-  const inflight = new Map()
+  const inflight = new Set()
+
+  ctx.effect(() => () => {
+    for (const controller of inflight) controller.abort()
+    inflight.clear()
+    cache.clear()
+  })
 
   ctx.effect(() =>
     ctx.webServer.register({
@@ -206,17 +212,30 @@ function describeError(error) {
  * @returns `null` when the response was already written, a plan, or an error.
  */
 async function planRequest(ctx, request, response) {
+  const admission = ctx.connection.admit(request)
+  if ('rejection' in admission) {
+    return { status: admission.rejection, error: admission.rejection === 401 ? 'unauthorized' : 'forbidden' }
+  }
+  // Predictions read the owner's sessions and spend the owner's model tokens.
+  if (admission.peer.principal.kind !== 'owner') return { status: 403, error: 'forbidden' }
+
   if (request.method !== 'POST') {
     respond(response, 405, { suggestion: null, suggestions: [], reason: 'method-not-allowed' })
     return null
   }
 
+  if (request.headers['content-type']?.split(';', 1)[0].trim().toLowerCase() !== 'application/json') {
+    return { status: 415, error: 'unsupported-media-type' }
+  }
+
   let body
   try {
     body = await readJsonBody(request)
   } catch {
     return { status: 400, error: 'invalid-body' }
   }
+  if (body === null || typeof body !== 'object' || Array.isArray(body)) return { status: 400, error: 'invalid-body' }
+  if (body.mode !== undefined && body.mode !== 'next' && body.mode !== 'continue') return { status: 400, error: 'invalid-mode' }
 
   const sessionId = typeof body?.sessionId === 'string' ? body.sessionId : ''
   if (sessionId === '') return { status: 400, error: 'missing-session' }
@@ -229,10 +248,7 @@ async function planRequest(ctx, request, response) {
   if (mode === 'continue' && draft.trim() === '') return { status: 200, error: 'no-draft' }
 
   const seq = Number(session.seq)
-  const key =
-    mode === 'next'
-      ? `${sessionId}:${seq}`
-      : `${sessionId}:${seq}:d${hashText(draft)}`
+  const key = JSON.stringify([sessionId, seq, mode, draft])
 
   return { sessionId, session, mode, draft, seq, key }
 }
@@ -247,17 +263,17 @@ async function runPrediction(ctx, options, cache, inflight, plan, signals, onDel
     return { items: remembered.items, reason: undefined, cached: true }
   }
 
-  // Supersede only an identical request; another view or the JSON endpoint on
-  // the same Session must not cancel this one.
-  inflight.get(plan.key)?.abort()
+  // Each view owns its cancellation; an identical request must not cancel it.
   const controller = new AbortController()
-  inflight.set(plan.key, controller)
+  inflight.add(controller)
   try {
     const result = await predict(ctx, options, plan.session, plan.mode, plan.draft, joinSignals([controller.signal, ...signals]), onDelta)
-    if (result.items.length > 0) remember(cache, plan.key, { items: result.items })
+    if (result.items.length > 0 && !controller.signal.aborted && !signals.some(signal => signal.aborted)) {
+      remember(cache, plan.key, { items: result.items })
+    }
     return result
   } finally {
-    if (inflight.get(plan.key) === controller) inflight.delete(plan.key)
+    inflight.delete(controller)
   }
 }
 
@@ -280,6 +296,7 @@ async function predict(ctx, options, session, mode, draft, signal, onDelta) {
   let text = ''
   let finish
   try {
+    deadline.signal.throwIfAborted()
     const call = await ctx.llm.prepareCall(
       { provider: route.provider, model: route.model, maxTokens: options.maxOutputTokens },
       deadline.signal,
@@ -290,6 +307,7 @@ async function predict(ctx, options, session, mode, draft, signal, onDelta) {
     )
     const messages = [{ role: 'user', content: [{ type: 'text', text: prompt }] }]
     for await (const chunk of call.stream({ ...call.config, messages, system, signal: deadline.signal })) {
+      deadline.signal.throwIfAborted()
       if (chunk.type === 'text-delta') {
         text += chunk.text
         onDelta?.(chunk.text)
@@ -297,6 +315,7 @@ async function predict(ctx, options, session, mode, draft, signal, onDelta) {
         finish = chunk.reason
       }
     }
+    deadline.signal.throwIfAborted()
   } finally {
     deadline.dispose()
   }
@@ -305,6 +324,7 @@ async function predict(ctx, options, session, mode, draft, signal, onDelta) {
   // outcome reports its own failure instead of looking like a missing stream.
   const failure = finishFailure(finish)
   if (failure !== undefined) return { items: [], reason: failure }
+  if (finish.kind === 'max-tokens') text = text.slice(0, text.lastIndexOf('\n') + 1)
 
   const cue = mode === 'continue' ? CONTINUE_CUE : NEXT_CUE
   const items = normalizeCandidates(text, options.maxSuggestions, mode === 'continue' ? draft : '', cue)
@@ -519,16 +539,6 @@ function boundSuggestion(line) {
   return space > 120 ? cut.slice(0, space) : cut
 }
 
-/** FNV-1a over the draft: a stable, cheap cache identity. */
-function hashText(text) {
-  let hash = 0x811c9dc5
-  for (let index = 0; index < text.length; index += 1) {
-    hash ^= text.charCodeAt(index)
-    hash = Math.imul(hash, 0x01000193)
-  }
-  return (hash >>> 0).toString(36)
-}
-
 function readJsonBody(request) {
   return new Promise((resolve, reject) => {
     const chunks = []

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-composer-predictions",
-  "version": "1.2.0",
-  "description": "Tab-completion for the DeepSeek Harness composer: next-message ghost text plus inline continuation of your draft, with keyboard candidate cycling",
+  "version": "1.2.2",
+  "description": "Tab autocomplete for the DeepSeek Harness desktop and web composer: next-message ghost text, inline draft completion and keyboard candidate cycling",
   "type": "module",
   "license": "MIT",
   "scripts": {
@@ -14,7 +14,12 @@
     "deepseek-harness",
     "composer",
     "autocomplete",
-    "tab-completion"
+    "tab-completion",
+    "ghost-text",
+    "inline-completion",
+    "prediction",
+    "desktop",
+    "input-completion"
   ],
   "publishConfig": {
     "access": "public"
@@ -53,5 +58,8 @@
   "homepage": "https://github.com/Stijnus/dsh-composer-predictions#readme",
   "bugs": {
     "url": "https://github.com/Stijnus/dsh-composer-predictions/issues"
+  },
+  "engines": {
+    "dsh": ">=0.2.1-rc.3 <0.3.0-0"
   }
 }