dsh-cline-pass 0.2.11 → 0.2.12

package/lib/adapter.js

@@ -54,17 +54,7 @@ export const DEFAULT_MAX_CONCURRENT_REQUESTS = 16
  * unknown ones; raise it per model in configuration when a new model ships.
  */
 export const DEFAULT_CONTEXT_WINDOW = 128000
-/**
- * Per-request output budget for a model whose caller names no `maxTokens`.
- *
- * A budget, not a ceiling: the harness materializes it into requests that omit
- * `maxTokens` and reserves the same number as completion room before it compacts,
- * so this is the number that decides how much of the window stays available to the
- * conversation. 64k is far past any single response a tool-calling agent emits
- * while keeping the reserve well under the point where it would start displacing
- * the compaction ratio. The published ceiling clamps it per model.
- */
-export const DEFAULT_MAX_TOKENS = 64000
+export { DEFAULT_MAX_TOKENS } from './catalog.js'
 /**
  * Request-image budget applied when the caller names none.
  *

package/lib/catalog.js

@@ -167,6 +167,22 @@ function pick(configured, published, fallback) {
   return fallback
 }
 
+/**
+ * Per-request output budget used when nothing else names one.
+ *
+ * A budget, not a ceiling: the harness materializes it into requests that omit
+ * `maxTokens`, and reserves the same number as completion room before it compacts, so
+ * this is the number that decides how much of the window stays available to the
+ * conversation.
+ *
+ * 128k doubles the room a long reasoning pass gets without moving that point on a 1M
+ * window: the 80% rule stays the binding one until the reserve passes ~144k. It is
+ * not applied flat, though — a 128k reserve on a 256k window would leave almost
+ * nothing to compact into — so the shipped default scales to the window while a
+ * configured value does not, because that one is a choice rather than a default.
+ */
+export const DEFAULT_MAX_TOKENS = 131072
+
 /**
  * The seam's whole `ModelModality` vocabulary. A catalog entry may list more
  * than this (models.dev reports `audio` for mimo-v2.5); the seam would reject
@@ -219,8 +235,11 @@ function resolveOutputBudget(budget, ceiling, contextWindow) {
   const most = Number.isFinite(ceiling) && ceiling > 0 ? ceiling : undefined
   const chosen = wanted === undefined ? most : most === undefined ? wanted : Math.min(wanted, most)
   if (chosen === undefined) return undefined
-  const half = Number.isFinite(contextWindow) && contextWindow > 0 ? Math.floor(contextWindow / 2) : undefined
-  return half === undefined ? chosen : Math.min(chosen, half)
+  if (!Number.isFinite(contextWindow) || contextWindow <= 0) return chosen
+  // The shipped default follows the window; a configured value is the user's choice
+  // and is only held back from reserving the window outright.
+  const share = wanted === DEFAULT_MAX_TOKENS ? 8 : 2
+  return Math.min(chosen, Math.floor(contextWindow / share))
 }
 
 /**

package/lib/client.js

@@ -77,6 +77,7 @@ window.__ModuleLoader__.load({
       '.cp-input,.cp-select{background:var(--dsw-alias-bg-base);color:var(--dsw-alias-label-primary);border:0.5px solid var(--dsw-alias-border-l2);border-radius:6px;padding:6px 8px;font-size:13px;min-width:0}',
       '.cp-input:focus-visible,.cp-select:focus-visible{outline:2px solid var(--dsw-alias-brand-primary);outline-offset:1px}',
       '.cp-btn{border:0.5px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-primary);border-radius:6px;padding:5px 10px;font-size:12px;cursor:pointer}',
+      '.cp-btn-on{border-color:var(--dsw-alias-brand-primary);color:var(--dsw-alias-brand-primary)}',
       '.cp-btn:hover:not(:disabled){border-color:var(--dsw-alias-brand-primary)}',
       '.cp-btn:disabled{opacity:.5;cursor:default}',
       '.cp-btn-primary{background:var(--dsw-alias-brand-primary);border-color:var(--dsw-alias-brand-primary);color:var(--dsw-alias-bg-base);font-weight:600}',
@@ -199,6 +200,7 @@ window.__ModuleLoader__.load({
       '.cp-history-load-row>span{white-space:nowrap}',
       '.cp-history-load-row>span:first-child{flex:none}',
       '.cp-history-load-row>span:last-child{min-width:0;overflow:hidden;text-overflow:ellipsis;text-align:right}',
+      '.cp-history-token-stack{display:flex;flex-direction:column;gap:1px;align-items:center}',
       '.cp-history-meta{display:flex;align-items:center;justify-content:center;gap:6px;min-width:0;overflow:hidden}',
       // A failure gets its own row under the entry rather than widening it: the
       // gateway's messages are long enough to wreck a one-line layout.
@@ -360,6 +362,7 @@ window.__ModuleLoader__.load({
       historyUpstream: '上游',
       historyLoad: '延迟',
       historyTokens: 'TOKEN',
+      historyCacheHit: '缓存',
       historyFirstChunk: '首字',
       historyTotal: '总耗时',
       historyRate: '输出速率',
@@ -381,6 +384,11 @@ window.__ModuleLoader__.load({
       refresh: '刷新',
       channels: '渠道（点选＝钉住顺序，⊘＝排除）',
       channelPlaceholder: '输入渠道名，回车添加',
+      budget: '输出预算',
+      budgetCurrent: '{n} tok',
+      budgetRecommended: '128K（推荐）',
+      budgetOfficial: '256K（对齐官方）',
+      budgetCustom: '自定义 token 数',
       channelAdd: '添加',
       channelUnverified: '未验证',
       channelsKnown: '已探测：{list}',
@@ -466,6 +474,7 @@ window.__ModuleLoader__.load({
       historyUpstream: 'Upstream',
       historyLoad: 'Latency',
       historyTokens: 'Tokens',
+      historyCacheHit: 'cache',
       historyFirstChunk: 'first',
       historyTotal: 'total',
       historyRate: 'rate',
@@ -487,6 +496,11 @@ window.__ModuleLoader__.load({
       refresh: 'Refresh',
       channels: 'Channels (click = pin order, ⊘ = exclude)',
       channelPlaceholder: 'Channel name, Enter to add',
+      budget: 'Output budget',
+      budgetCurrent: '{n} tok',
+      budgetRecommended: '128K (recommended)',
+      budgetOfficial: '256K (matches official)',
+      budgetCustom: 'Custom token count',
       channelAdd: 'Add',
       channelUnverified: 'unverified',
       channelsKnown: 'Probed: {list}',
@@ -784,6 +798,7 @@ window.__ModuleLoader__.load({
           pinModel: (payload) => controller.pinModel(payload),
           setModelVisible: (model, visible) => controller.setModelVisible(model, visible),
           setToolsExposed: (value) => controller.setToolsExposed(value),
+          setBudget: (value) => controller.setBudget(value),
           setModelsVisibility: (mode) => controller.setModelsVisibility(mode),
           probeModel: (model) => controller.probeModel(model),
           validateModel: (model) => controller.validateModel(model),
@@ -974,6 +989,13 @@ window.__ModuleLoader__.load({
         return result
       }
 
+      /** Set the per-request output budget, in tokens. */
+      async setBudget(value) {
+        const result = await this.run('budget', async () => await this.call('budget.set', { value }))
+        if (result !== undefined) this.applyState(result)
+        return result
+      }
+
       /**
        * Show, hide or swap every subscription model at once.
        *
@@ -1601,6 +1623,60 @@ window.__ModuleLoader__.load({
       ])
     }
 
+    /**
+     * The per-request output budget.
+     *
+     * One number for the whole route: the harness reserves it as completion room on
+     * every request, so it decides where compaction fires as much as how long one
+     * response may be. The two presets are the shipped default and the figure
+     * DeepSeek's own provider uses; anything else is typed.
+     */
+    function BudgetCard(props) {
+      const { t, data, disabled, actions } = props
+      const current = Number(data.defaultMaxTokens ?? 0)
+      const [draft, setDraft] = React.useState('')
+      const write = (value) => {
+        if (disabled) return
+        actions.setBudget(value)
+      }
+      const custom = () => {
+        const value = Number(draft.trim())
+        if (!Number.isSafeInteger(value) || value < 1024) return
+        setDraft('')
+        write(value)
+      }
+      return h('div', { className: 'cp-card' }, [
+        h('div', { key: 'title', className: 'cp-card-title' }, `${t('budget')} · ${fill(t('budgetCurrent'), { n: current.toLocaleString('en-US') })}`),
+        h('div', { key: 'row', className: 'cp-row' }, [
+          h('button', {
+            key: 'recommended', type: 'button',
+            className: `cp-btn${current === 131072 ? ' cp-btn-on' : ''}`,
+            disabled, onClick: () => write(131072),
+          }, t('budgetRecommended')),
+          h('button', {
+            key: 'official', type: 'button',
+            className: `cp-btn${current === 256000 ? ' cp-btn-on' : ''}`,
+            disabled, onClick: () => write(256000),
+          }, t('budgetOfficial')),
+        ]),
+        h('div', { key: 'custom', className: 'cp-row' }, [
+          h('input', {
+            key: 'in', type: 'text', className: 'cp-input cp-grow',
+            value: draft, disabled,
+            placeholder: t('budgetCustom'),
+            'aria-label': t('budget'),
+            onChange: (event) => setDraft(event.target.value),
+            onKeyDown: (event) => {
+              if (event.key !== 'Enter') return
+              event.preventDefault()
+              custom()
+            },
+          }),
+          h('button', { key: 'b', type: 'button', className: 'cp-btn', disabled, onClick: custom }, t('channelAdd')),
+        ]),
+      ])
+    }
+
     /**
      * The models card: a fold-away section, like the request history.
      *
@@ -1666,6 +1742,7 @@ window.__ModuleLoader__.load({
       // — centred text that overflows loses BOTH ends, not just the tail.
       const label = String(entry.model).replace(/^cline-pass\//, '')
       const figures = latency(entry)
+      const cacheHit = entry.cacheHit === undefined ? null : entry.cacheHit
       const row = h('tr', {
         key: 'row',
         // The whole row restates itself on hover, including the figures the three
@@ -1699,8 +1776,11 @@ window.__ModuleLoader__.load({
         h('td', {
           key: 'tokens',
           className: 'cp-history-tokens',
-          title: `${tokenSummaryExact(entry)}\n${t('historyTokensHint')}`,
-        }, tokenSummary(entry)),
+          title: `${tokenSummaryExact(entry)}\n${cacheHit === null ? '' : `${t('historyCacheHit')} ${cacheHit.toFixed(2)}%\n`}${t('historyTokensHint')}`,
+        }, h('div', { className: 'cp-history-token-stack' }, [
+          h('span', { key: 'sum' }, tokenSummary(entry)),
+          cacheHit === null ? null : h('span', { key: 'cache', className: 'cp-muted cp-small' }, `${t('historyCacheHit')} ${cacheHit.toFixed(2)}%`),
+        ])),
         h('td', { key: 'load', className: 'cp-history-load' }, [
           h('span', { key: 'first', className: 'cp-history-load-row' }, [
             h('span', { key: 'k', className: 'cp-muted' }, t('historyFirstChunk')),
@@ -1953,6 +2033,7 @@ window.__ModuleLoader__.load({
         pinModel: props.pinModel,
         setModelVisible: props.setModelVisible,
         setToolsExposed: props.setToolsExposed,
+        setBudget: props.setBudget,
         setModelsVisibility: props.setModelsVisibility,
         probeModel: props.probeModel,
         validateModel: props.validateModel,
@@ -2010,8 +2091,9 @@ window.__ModuleLoader__.load({
         // Everything below the accounts card reads through the key that card
         // stores, so none of it exists until there is one.
         data === null || !ready ? null : h(ModelsCard, { key: 'models', t, data, disabled, busy: state.busy, actions }),
-        // Route-wide switches last: they are not about one model or account, and
+        // Route-wide settings last: they are not about one model or account, and
         // nothing below them depends on their value.
+        data === null ? null : h(BudgetCard, { key: 'budget', t, data, disabled: state.busy !== null || data.settingsAvailable === false, actions }),
         data === null ? null : h('div', { key: 'tools', className: 'cp-card' }, [
           h('label', { key: 'head', className: 'cp-row cp-tools-label' }, [
             h('input', {

package/lib/index.js

@@ -160,6 +160,11 @@ export const Config = z.object({
    * larger, and reporting that here reserves it on every request — at the shipped
    * 384k figure a 1M window compacts at 55% instead of the 80% the threshold ratio
    * asks for. Each model's published ceiling clamps whatever is set here.
+   *
+   * Nothing here may reserve more than half the window. The shipped default goes
+   * further and scales to an eighth on a small window, because a flat 128k reserve
+   * leaves a 256k model almost nothing to compact into; an explicit value counts as a
+   * choice and only gets the half-window limit.
    */
   defaultMaxTokens: z.number().default(DEFAULT_MAX_TOKENS),
   streamIdleTimeoutMs: z.number().default(DEFAULT_STREAM_IDLE_TIMEOUT_MS),

package/lib/panel.js

@@ -130,6 +130,9 @@ export function createPanel({ control, engine, store }) {
       provider: control.providerName,
       displayName: control.displayName,
       baseURL: String(config.baseURL ?? ''),
+      // The per-request output budget: one number for the whole route, because it
+      // decides how much of every window the harness holds back as completion room.
+      defaultMaxTokens: Number(config.defaultMaxTokens ?? 0),
       settingsAvailable: control.settingsAvailable(),
       accountMode: config.accountMode === 'roundrobin' ? 'roundrobin' : 'single',
       activeAccount: String(config.activeAccount ?? ''),
@@ -449,6 +452,21 @@ export function createPanel({ control, engine, store }) {
     return await modelPin({ model: modelOf(payload), upstreams: [], exclude: [], sort: 'none', pinMode: 'preferred' })
   }
 
+  /**
+   * Cache reads as a share of billed prompt tokens, or null when nothing was measured.
+   *
+   * A ratio over zero is not a reading, and 0% would claim "nothing cached" for a
+   * request the gateway never reported usage for.
+   */
+  function cacheHitOf(usage) {
+    if (usage === null || usage === undefined) return null
+    const cached = Number(usage.cacheReadTokens ?? 0)
+    const billed = cached + Number(usage.inputTokens ?? 0)
+    // Two decimals: an integer round turns 99.6% into a flat 100%, which reads as a
+    // claim the numbers do not support.
+    return billed <= 0 ? null : Math.round((cached / billed) * 10000) / 100
+  }
+
   /** The live request history, newest first. */
   async function history(payload) {
     const limit = Number.isSafeInteger(payload?.limit) && payload.limit > 0 ? Math.min(payload.limit, 100) : 25
@@ -473,6 +491,7 @@ export function createPanel({ control, engine, store }) {
           cacheReadTokens: Number(entry.usage?.cacheReadTokens ?? 0),
           reasoningTokens: Number(entry.usage?.reasoningTokens ?? 0),
         },
+        cacheHit: cacheHitOf(entry.usage),
         effort: String(entry.effort ?? ''),
         stream: entry.stream === true,
         error: firstLine(entry.error, 200),
@@ -490,6 +509,23 @@ export function createPanel({ control, engine, store }) {
    * The response carries `restartRequired` so the panel can say so instead of
    * implying the switch already took effect.
    */
+  /**
+   * Set the per-request output budget.
+   *
+   * The panel offers the two numbers that matter — the shipped default, and the one
+   * DeepSeek's own provider uses — plus whatever a user types. It is a route-wide
+   * setting rather than a per-model one because it is the same policy for every
+   * model: how much of a window to hold back for one response.
+   */
+  async function budgetSet(payload) {
+    const value = Number(payload?.value)
+    if (!Number.isSafeInteger(value) || value < 1024 || value > 2000000) {
+      throw new Error(`the output budget must be a whole number of tokens between 1024 and 2000000 (got ${JSON.stringify(payload?.value)})`)
+    }
+    await control.updateConfig({ defaultMaxTokens: value })
+    return await state()
+  }
+
   async function toolsExpose(payload) {
     if (typeof payload?.value !== 'boolean') throw new Error('this action needs a boolean `value`')
     await control.updateConfig({ exposeTools: payload.value })
@@ -514,6 +550,7 @@ export function createPanel({ control, engine, store }) {
     'model.reset': modelReset,
     'models.refresh': modelsRefresh,
     'tools.expose': toolsExpose,
+    'budget.set': budgetSet,
     history,
   }
 }

package/lib/protocol.js

@@ -59,7 +59,15 @@ export function isAuthenticationError(message, status) {
 export function parseRouting(json) {
   const payload = unwrapEnvelope(json)
   const message = payload?.choices?.[0]?.message
-  const routing = message?.provider_metadata?.gateway?.routing ?? payload?.provider_metadata?.gateway?.routing ?? {}
+  // A non-streaming answer carries routing on the message; a stream carries the same
+  // object on the delta instead. Reading only the message left every streamed request
+  // with no upstream at all, which the history then filled in with the pinned name —
+  // reporting a pin the gateway may never have adopted.
+  const delta = payload?.choices?.[0]?.delta
+  const routing = message?.provider_metadata?.gateway?.routing
+    ?? delta?.provider_metadata?.gateway?.routing
+    ?? payload?.provider_metadata?.gateway?.routing
+    ?? {}
   const direct = typeof payload?.provider === 'string' ? payload.provider : null
   return {
     content: typeof message?.content === 'string' ? message.content : null,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-cline-pass",
-  "version": "0.2.11",
+  "version": "0.2.12",
   "description": "Self-contained Cline Pass provider for dsh: register Cline Pass subscription models on the harness LLM seam, enumerate and validate the upstream channels behind each model, pin or exclude them with failover, and drive an account pool. No proxy process and no external cline-pass-switcher install required.",
   "type": "module",
   "main": "lib/index.js",