dsh-vision-router 2.2.1 → 2.2.3

package/lib/client.js

@@ -71,9 +71,14 @@ window.__ModuleLoader__.load({
       localOllamaModel: 'Ollama 模型名',
       localOllamaTemperature: '温度 temperature',
       localOllamaTopP: 'top_p',
+      localMaxTokens: '最大输出 token',
+      localReasoningEffort: '推理强度',
+      localReasoningDefault: '服务端默认',
+      localReasoningNone: '关闭推理',
       localRequestFormat: '请求协议',
       localFormatOpenAI: 'OpenAI（/chat/completions）',
       localFormatAnthropic: 'Anthropic（/messages）',
+      localFormatLmStudio: 'LM Studio 原生（/api/v1/chat）',
       localTemperaturePlaceholder: '留空=服务端默认；建议 0.5',
       localTopPPlaceholder: '留空=服务端默认；建议 0.8',
       localLmStudioModelPlaceholder: '填写 Developer 页或 /v1/models 中的模型标识',
@@ -219,7 +224,7 @@ window.__ModuleLoader__.load({
         '用于兼容旧版行为，一般无需开启；开启后只使用上方识图模型链中的后端。',
       hintReverseRouting: '仅在「整轮交给视觉模型」开启时生效；纯文字消息继续交给聊天模型处理。',
       hintTool: '允许聊天模型按需查看、定位、裁剪和比较图片。推荐保持开启。',
-      hintStructuredVisionBootstrap: '默认关闭。开启后，每个图片任务会先做一次不读取具体任务目标的结构化预识别，再由聊天模型根据原问题至少追加一次验证或深挖识图调用（1+x，x≥1）。启用的 Ollama / LM Studio 会和其他视觉后端一样参与这条识图链，不需要另开「即时识图」。准确性更高，但会增加至少一次视觉调用；若后续需要 OCR，engine=auto 仍会先尝试本地 Tesseract，失败或空结果再回退视觉模型；结构化模式不会改变这一顺序。需保持「识图工具」开启。',
+      hintStructuredVisionBootstrap: '默认关闭。开启后，每个图片任务会先做一次不读取具体任务目标的结构化预识别，再由聊天模型根据原问题至少追加一次验证或深挖识图调用（1+x，x≥1）。启用的 Ollama / LM Studio 会和其他视觉后端一样参与这条识图链，不需要另开「即时识图」。准确性更高，但会增加至少一次视觉调用；若后续需要 OCR，未显式指定 engine 时会遵循「OCR 默认引擎」设置；单次显式 engine=tesseract/vision 始终优先。需保持「识图工具」开启。',
       hintAutoWrapProviders: '自动为已启用的聊天模型创建带「+ 自动识图」的版本。原模型不受影响，推荐保持开启。',
       hintRewriteImages: '避免把无法读取的原始图片直接发送给纯文字模型。推荐保持开启。',
       hintDownscale: '超过像素预算的图片先缩放再送视觉模型，降低延迟与成本；默认开启。',
@@ -243,6 +248,11 @@ window.__ModuleLoader__.load({
       numTimeoutMs: '单次识图请求超时（毫秒）',
       numVisionTaskTimeoutMs: '识图任务最长时间（毫秒）',
       numOcrTimeoutMs: 'OCR 最长时间（毫秒）',
+      selectOcrEngine: 'OCR 默认引擎',
+      hintOcrEngine: '决定 vision_ocr 未显式指定 engine 时的默认策略。自动：先本地 Tesseract，失败/空结果再用视觉模型；Tesseract：只用本地 OCR；视觉模型：直接跳过 Tesseract。单次工具调用显式 engine=tesseract/vision 始终优先。',
+      ocrEngineAuto: '自动（Tesseract → 视觉模型）',
+      ocrEngineTesseract: '仅 Tesseract',
+      ocrEngineVision: '仅视觉模型',
       numDownscaleMaxPixels: '图片像素上限',
       numCacheTtlSeconds: '缓存有效期（秒）',
       numCacheMaxEntries: '最大缓存数量',
@@ -347,9 +357,14 @@ window.__ModuleLoader__.load({
       localOllamaModel: 'Ollama model name',
       localOllamaTemperature: 'Temperature',
       localOllamaTopP: 'Top-p',
+      localMaxTokens: 'Max output tokens',
+      localReasoningEffort: 'Reasoning effort',
+      localReasoningDefault: 'Provider default',
+      localReasoningNone: 'Disable reasoning',
       localRequestFormat: 'Request protocol',
       localFormatOpenAI: 'OpenAI (/chat/completions)',
       localFormatAnthropic: 'Anthropic (/messages)',
+      localFormatLmStudio: 'LM Studio native (/api/v1/chat)',
       localTemperaturePlaceholder: 'Blank = server default; suggested 0.5',
       localTopPPlaceholder: 'Blank = server default; suggested 0.8',
       localLmStudioModelPlaceholder: 'Model identifier from Developer or /v1/models',
@@ -496,7 +511,7 @@ window.__ModuleLoader__.load({
         'only the vision models configured above participate.',
       hintReverseRouting: 'Only applies when whole-turn vision routing is enabled; plain text messages continue to use the chat model.',
       hintTool: 'Lets the chat model inspect, locate, crop, and compare images as needed. Recommended on.',
-      hintStructuredVisionBootstrap: 'Off by default. Each image task first gets one task-independent structured visual baseline with no task goal passed into the pre-scan, then the chat model must make at least one evidence or deepening vision call for the original request (1+x, x>=1). Enabled Ollama / LM Studio backends participate in this same vision chain; there is no separate instant-recognition switch. This improves evidence quality but adds at least one visual call. If OCR is needed, engine=auto still tries local Tesseract first and falls back to the vision model only when local OCR fails or returns no text; structured mode does not change this order. Keep Vision tools enabled.',
+      hintStructuredVisionBootstrap: 'Off by default. Each image task first gets one task-independent structured visual baseline with no task goal passed into the pre-scan, then the chat model must make at least one evidence or deepening vision call for the original request (1+x, x>=1). Enabled Ollama / LM Studio backends participate in this same vision chain; there is no separate instant-recognition switch. This improves evidence quality but adds at least one visual call. If OCR is needed, calls without an explicit engine follow the “Default OCR engine” setting; an explicit engine=tesseract/vision always wins for that call. Keep Vision tools enabled.',
       hintAutoWrapProviders: 'Automatically creates a “+ Auto Vision” version of enabled chat models. Original model groups stay unchanged. Recommended on.',
       hintRewriteImages: 'Prevents raw image content from being sent to a text-only model that cannot read it. Recommended on.',
       hintDownscale: 'Images beyond the pixel budget are resized before the vision call, cutting latency and cost; on by default.',
@@ -521,6 +536,11 @@ window.__ModuleLoader__.load({
       numTimeoutMs: 'Single vision request timeout (ms)',
       numVisionTaskTimeoutMs: 'Vision task maximum time (ms)',
       numOcrTimeoutMs: 'OCR maximum time (ms)',
+      selectOcrEngine: 'Default OCR engine',
+      hintOcrEngine: 'Default policy when vision_ocr does not explicitly choose an engine. Auto tries local Tesseract first and falls back to the vision model; Tesseract uses local OCR only; Vision skips Tesseract. An explicit engine=tesseract/vision on a tool call always wins.',
+      ocrEngineAuto: 'Auto (Tesseract → vision)',
+      ocrEngineTesseract: 'Tesseract only',
+      ocrEngineVision: 'Vision model only',
       numDownscaleMaxPixels: 'Image pixel limit',
       numCacheTtlSeconds: 'Cache TTL (seconds)',
       numCacheMaxEntries: 'Maximum cached answers',
@@ -719,10 +739,14 @@ window.__ModuleLoader__.load({
       localOllama: {
         baseURL: 'http://127.0.0.1:11434/v1',
         model: 'qwen2.5vl',
+        maxTokens: 4096,
+        reasoningEffort: 'none',
       },
       localLmStudio: {
         baseURL: 'http://localhost:1234/v1',
         model: '',
+        maxTokens: 4096,
+        reasoningEffort: 'none',
       },
     }
 
@@ -739,7 +763,17 @@ window.__ModuleLoader__.load({
           typeof input.model === 'string' && input.model !== ''
             ? input.model
             : defaults.model,
-        format: input.format === 'anthropic' ? 'anthropic' : 'openai',
+        format: input.format === 'anthropic' ? 'anthropic' : input.format === 'lmstudio' ? 'lmstudio' : 'openai',
+        ...((Number.isInteger(input.maxTokens) ? input.maxTokens : defaults.maxTokens) === undefined
+          ? {}
+          : { maxTokens: Number.isInteger(input.maxTokens) ? input.maxTokens : defaults.maxTokens }),
+        ...((['provider_default', 'none', 'low', 'medium', 'high', 'max'].includes(input.reasoningEffort)
+          ? input.reasoningEffort
+          : defaults.reasoningEffort) === undefined
+          ? {}
+          : { reasoningEffort: ['provider_default', 'none', 'low', 'medium', 'high', 'max'].includes(input.reasoningEffort)
+              ? input.reasoningEffort
+              : defaults.reasoningEffort }),
         temperature:
           typeof input.temperature === 'number' && Number.isFinite(input.temperature)
             ? input.temperature
@@ -761,6 +795,12 @@ window.__ModuleLoader__.load({
         typeof input.top_p === 'number' && Number.isFinite(input.top_p)
           ? Math.min(1, Math.max(0, input.top_p))
           : undefined
+      const maxTokens = Number.isInteger(input.maxTokens)
+        ? Math.min(32768, Math.max(256, input.maxTokens))
+        : defaults.maxTokens
+      const reasoningEffort = ['provider_default', 'none', 'low', 'medium', 'high', 'max'].includes(input.reasoningEffort)
+        ? input.reasoningEffort
+        : defaults.reasoningEffort
       return {
         enabled: input.enabled === true,
         baseURL:
@@ -771,7 +811,9 @@ window.__ModuleLoader__.load({
           typeof input.model === 'string' && input.model.trim() !== ''
             ? input.model.trim()
             : defaults.model,
-        format: input.format === 'anthropic' ? 'anthropic' : 'openai',
+        format: input.format === 'anthropic' ? 'anthropic' : input.format === 'lmstudio' ? 'lmstudio' : 'openai',
+        ...(maxTokens === undefined ? {} : { maxTokens }),
+        ...(reasoningEffort === undefined ? {} : { reasoningEffort }),
         ...(temperature === undefined ? {} : { temperature }),
         ...(topP === undefined ? {} : { top_p: topP }),
       }
@@ -865,6 +907,57 @@ window.__ModuleLoader__.load({
           : stored && settingsValueEqual(item.key, user[item.key], item.run.value)
         return { ok, stored }
       }
+      const finalize = () => {
+        const landed = failures.length === 0
+        const nextDrafts = landedFields.length === 0 ? drafts : { ...drafts }
+        for (const field of landedFields) delete nextDrafts[field]
+        return {
+          landed,
+          failed: !landed,
+          landedFields,
+          nextDrafts,
+          failures,
+        }
+      }
+      const mismatchFailure = (item, check) => ({
+        field: item.key,
+        operation: item.run.clear ? 'unset' : 'set',
+        reason: 'readback-mismatch',
+        detail: item.run.clear
+          ? 'field remained present in the user layer'
+          : check.stored
+            ? 'stored user-layer value differs from the requested value'
+            : 'field is absent from the user layer',
+      })
+      const batchWrite = scope && typeof scope.__visionRouterWritePlan === 'function'
+        ? scope.__visionRouterWritePlan.bind(scope)
+        : undefined
+      // DSH 0.1.7's local compatibility transport is ConfigEditor-backed.
+      // ConfigEditor reloads the plugin after a successful edit, so serial POSTs
+      // from one Save can cross generations and strand the second mutation.
+      // Its private scope seam commits the entire UI plan in one atomic edit.
+      if (plan.length > 1 && batchWrite) {
+        try {
+          await batchWrite(plan)
+        } catch (error) {
+          for (const item of plan) {
+            failures.push({
+              field: item.key,
+              operation: item.run.clear ? 'unset' : 'set',
+              reason: error && error.code === 'settings-conflict' ? 'settings-conflict' : 'write-error',
+              detail: settingsSaveErrorMessage(error),
+            })
+          }
+          return finalize()
+        }
+        for (const item of plan) {
+          const check = inspectReadback(item)
+          if (check.error) failures.push(check.error)
+          else if (check.ok) landedFields.push(item.key)
+          else failures.push(mismatchFailure(item, check))
+        }
+        return finalize()
+      }
       const writeItem = async (item) => {
         if (item.run.clear) await scope.unset(item.key)
         else await scope.set(item.key, item.run.value)
@@ -913,16 +1006,7 @@ window.__ModuleLoader__.load({
             }
           }
 
-          terminalFailure = {
-            field: item.key,
-            operation,
-            reason: 'readback-mismatch',
-            detail: item.run.clear
-              ? 'field remained present in the user layer'
-              : check.stored
-                ? 'stored user-layer value differs from the requested value'
-                : 'field is absent from the user layer',
-          }
+          terminalFailure = mismatchFailure(item, check)
         }
         if (success) landedFields.push(item.key)
         else failures.push(terminalFailure ?? {
@@ -932,16 +1016,7 @@ window.__ModuleLoader__.load({
           detail: 'write did not become visible in the user layer',
         })
       }
-      const landed = failures.length === 0
-      const nextDrafts = landedFields.length === 0 ? drafts : { ...drafts }
-      for (const field of landedFields) delete nextDrafts[field]
-      return {
-        landed,
-        failed: !landed,
-        landedFields,
-        nextDrafts,
-        failures,
-      }
+      return finalize()
     }
 
     const REMOTE_SETTINGS_CHANNEL = '/vision-router-settings'
@@ -1124,16 +1199,23 @@ window.__ModuleLoader__.load({
     }
 
     function shouldUseRemoteSettings(getConnection, locationLike) {
+      // Browser page authority is the security boundary. DSH 0.1.7 can report
+      // Connection.isLoopback=false even for a real 127.0.0.1 page, so letting
+      // that hint win would incorrectly force the local Settings page onto the
+      // trusted-host RPC path. Conversely, a non-loopback browser URL must not
+      // become local merely because a stale Connection says true.
+      const location = locationLike ?? (typeof window !== 'undefined' ? window.location : undefined)
+      const hostname = typeof location?.hostname === 'string' ? location.hostname.toLowerCase().replace(/^\[|\]$/g, '') : ''
+      if (hostname !== '') {
+        if (hostname === 'localhost' || hostname.endsWith('.localhost') || hostname === '::1') return false
+        if (/^127(?:\.\d{1,3}){3}$/.test(hostname)) return false
+        return true
+      }
       try {
         const connection = typeof getConnection === 'function' ? getConnection() : undefined
         if (connection && typeof connection.isLoopback === 'boolean') return connection.isLoopback === false
-      } catch { /* fall through to page authority */ }
-      const location = locationLike ?? (typeof window !== 'undefined' ? window.location : undefined)
-      const hostname = typeof location?.hostname === 'string' ? location.hostname.toLowerCase().replace(/^\[|\]$/g, '') : ''
-      if (hostname === '') return false
-      if (hostname === 'localhost' || hostname.endsWith('.localhost') || hostname === '::1') return false
-      if (/^127(?:\.\d{1,3}){3}$/.test(hostname)) return false
-      return true
+      } catch { /* no page authority and no usable Connection: stay local-safe */ }
+      return false
     }
 
     function reportSettingsSaveFailures(failures) {
@@ -2331,6 +2413,7 @@ window.__ModuleLoader__.load({
       timeoutMs: 'numTimeoutMs',
       visionTaskTimeoutMs: 'numVisionTaskTimeoutMs',
       ocrTimeoutMs: 'numOcrTimeoutMs',
+      ocrEngine: 'selectOcrEngine',
       downscaleMaxPixels: 'numDownscaleMaxPixels',
       cacheTtlSeconds: 'numCacheTtlSeconds',
       cacheMaxEntries: 'numCacheMaxEntries',
@@ -2359,6 +2442,7 @@ window.__ModuleLoader__.load({
       timeoutMs: 'numHintTimeoutMs',
       visionTaskTimeoutMs: 'numHintVisionTaskTimeoutMs',
       ocrTimeoutMs: 'numHintOcrTimeoutMs',
+      ocrEngine: 'hintOcrEngine',
       downscaleMaxPixels: 'numHintDownscaleMaxPixels',
       cacheTtlSeconds: 'numHintCacheTtlSeconds',
       cacheMaxEntries: 'numHintCacheMaxEntries',
@@ -2756,6 +2840,7 @@ window.__ModuleLoader__.load({
           return normalizeLocalProviderDraft(value, LOCAL_PROVIDER_DEFAULTS.localLmStudio)
         }
         if (key === 'localDescribeStyle') return value === 'structured' ? 'structured' : 'plain'
+        if (key === 'ocrEngine') return value === 'tesseract' || value === 'vision' ? value : 'auto'
         if (SELECT_KEYS.includes(key)) {
           if (value === 'custom') return 'standard'
           return value === 'fast' || value === 'standard' || value === 'deep' ? value : 'standard'
@@ -2816,6 +2901,9 @@ window.__ModuleLoader__.load({
         if (key === 'localDescribeStyle') {
           return text === 'structured' || text === 'plain' ? { value: text } : undefined
         }
+        if (key === 'ocrEngine') {
+          return text === 'auto' || text === 'tesseract' || text === 'vision' ? { value: text } : undefined
+        }
         if (SELECT_KEYS.includes(key)) {
           return text === 'fast' || text === 'standard' || text === 'deep' ? { value: text } : undefined
         }
@@ -3314,6 +3402,26 @@ window.__ModuleLoader__.load({
               h('option', { value: 'anthropic' }, t('localFormatAnthropic')),
             ),
           ),
+          h('div', { className: 'vr-local-row vr-local-row-pair' },
+            h('label', { className: 'vr-label vr-local-label' }, t('localMaxTokens')),
+            h('input', {
+              className: 'vr-input', value: String(value.maxTokens ?? 4096), disabled: editBlocked,
+              type: 'number', step: '256', min: '256', max: '32768',
+              onChange: (event) => update({ maxTokens: Number(event.target.value) }),
+            }),
+            h('label', { className: 'vr-label vr-local-label' }, t('localReasoningEffort')),
+            h('select', {
+              className: 'vr-input vr-select', value: value.reasoningEffort, disabled: editBlocked || value.format === 'anthropic',
+              onChange: (event) => update({ reasoningEffort: event.target.value }),
+            },
+              h('option', { value: 'provider_default' }, t('localReasoningDefault')),
+              h('option', { value: 'none' }, t('localReasoningNone')),
+              h('option', { value: 'low' }, 'low'),
+              h('option', { value: 'medium' }, 'medium'),
+              h('option', { value: 'high' }, 'high'),
+              h('option', { value: 'max' }, 'max'),
+            ),
+          ),
           h('div', { className: 'vr-local-row vr-local-row-pair' },
             h('label', { className: 'vr-label vr-local-label' }, t('localOllamaTemperature')),
             h('input', {
@@ -3378,6 +3486,27 @@ window.__ModuleLoader__.load({
             },
               h('option', { value: 'openai' }, t('localFormatOpenAI')),
               h('option', { value: 'anthropic' }, t('localFormatAnthropic')),
+              h('option', { value: 'lmstudio' }, t('localFormatLmStudio')),
+            ),
+          ),
+          h('div', { className: 'vr-local-row vr-local-row-pair' },
+            h('label', { className: 'vr-label vr-local-label' }, t('localMaxTokens')),
+            h('input', {
+              className: 'vr-input', value: String(value.maxTokens ?? 4096), disabled: editBlocked,
+              type: 'number', step: '256', min: '256', max: '32768',
+              onChange: (event) => update({ maxTokens: Number(event.target.value) }),
+            }),
+            h('label', { className: 'vr-label vr-local-label' }, t('localReasoningEffort')),
+            h('select', {
+              className: 'vr-input vr-select', value: value.reasoningEffort, disabled: editBlocked || value.format !== 'lmstudio',
+              onChange: (event) => update({ reasoningEffort: event.target.value }),
+            },
+              h('option', { value: 'provider_default' }, t('localReasoningDefault')),
+              h('option', { value: 'none' }, t('localReasoningNone')),
+              h('option', { value: 'low' }, 'low'),
+              h('option', { value: 'medium' }, 'medium'),
+              h('option', { value: 'high' }, 'high'),
+              h('option', { value: 'max' }, 'max'),
             ),
           ),
           h('div', { className: 'vr-local-row vr-local-row-pair' },
@@ -4083,6 +4212,11 @@ window.__ModuleLoader__.load({
             h('p', { className: 'vr-group-title' }, t('groupPerformance')),
             PERFORMANCE_TOGGLE_KEYS.map((key) => toggleField(key)),
             NUMBER_KEYS.map((key) => textField(key, t(LABEL_KEY[key]), t(HINT_KEY[key]), false)),
+            selectField('ocrEngine', t(LABEL_KEY.ocrEngine), t(HINT_KEY.ocrEngine), [
+              { value: 'auto', label: t('ocrEngineAuto') },
+              { value: 'tesseract', label: t('ocrEngineTesseract') },
+              { value: 'vision', label: t('ocrEngineVision') },
+            ]),
           ),
           h('div', { className: 'vr-group' },
             h('p', { className: 'vr-group-title' }, t('groupCompatibility')),

package/lib/core-primitives.js

@@ -1439,8 +1439,9 @@ export function posterizeSvgColor(data, info, palette, timeoutMs = 60000) {
 }
 
 /** Resolve the effective vision_ocr engine without hiding explicit user/model intent. */
-export function resolveVisionOcrEngine(requestedEngine) {
+export function resolveVisionOcrEngine(requestedEngine, configuredEngine = 'auto') {
   if (requestedEngine === 'tesseract' || requestedEngine === 'vision') return requestedEngine
+  if (configuredEngine === 'tesseract' || configuredEngine === 'vision') return configuredEngine
   return 'auto'
 }
 
@@ -1857,11 +1858,15 @@ export function localOllamaProvidersOf(config) {
       baseURL,
       model,
       apiKeyEnv: '',
-      maxTokens: 2048,
+      maxTokens: Number.isInteger(local.maxTokens) ? local.maxTokens : 4096,
       // Ollama's OpenAI-compatible Chat Completions endpoint supports
       // reasoning_effort='none'. Vision extraction should spend the bounded
       // completion budget on observable answer text rather than hidden thought.
-      reasoningEffort: 'none',
+      ...(local.reasoningEffort === 'provider_default'
+        ? {}
+        : { reasoningEffort: ['none', 'low', 'medium', 'high', 'max'].includes(local.reasoningEffort)
+            ? local.reasoningEffort
+            : 'none' }),
       // 仅显式选择 anthropic 格式时携带（默认 openai 路径保持字节不变）。
       ...(local.format === 'anthropic' ? { format: 'anthropic' } : {}),
       // 建议值透传：温度/top_p 只在显式配置时携带（callOpenAICompatible
@@ -1889,8 +1894,12 @@ export function localLmStudioProvidersOf(config) {
       baseURL,
       model,
       apiKeyEnv: '',
-      maxTokens: 2048,
+      maxTokens: Number.isInteger(local.maxTokens) ? local.maxTokens : 4096,
+      ...(local.format === 'lmstudio' && ['none', 'low', 'medium', 'high', 'max'].includes(local.reasoningEffort)
+        ? { reasoningEffort: local.reasoningEffort }
+        : {}),
       ...(local.format === 'anthropic' ? { format: 'anthropic' } : {}),
+      ...(local.format === 'lmstudio' ? { format: 'lmstudio' } : {}),
       ...(typeof local.temperature === 'number' ? { temperature: local.temperature } : {}),
       ...(typeof local.top_p === 'number' ? { top_p: local.top_p } : {}),
     },
@@ -1917,12 +1926,108 @@ export function localProvidersOf(config) {
  * temperature/top_p 仅显式配置时透传（两个 transport 的显式可选参数，
  * 现有调用不传，wire 保持 main 原样）。
  */
+function lmStudioNativeContent(messages) {
+  const system = []
+  const input = []
+  for (const message of messages ?? []) {
+    if (!message) continue
+    const blocks = Array.isArray(message.content)
+      ? message.content
+      : typeof message.content === 'string'
+        ? [{ type: 'text', text: message.content }]
+        : []
+    if (message.role === 'system') {
+      for (const block of blocks) {
+        if (block?.type === 'text' && typeof block.text === 'string' && block.text.trim() !== '') {
+          system.push(block.text)
+        }
+      }
+      continue
+    }
+    for (const block of blocks) {
+      if (block?.type === 'text' && typeof block.text === 'string') {
+        input.push({ type: 'text', content: block.text })
+      } else if (block?.type === 'image_url' && typeof block.image_url?.url === 'string') {
+        input.push({ type: 'image', data_url: block.image_url.url })
+      }
+    }
+  }
+  return { systemPrompt: system.join('\n').trim(), input }
+}
+
+async function callLmStudioNative(provider, messages, options = {}) {
+  const normalizedBaseURL = stripTrailingSlashes(String(provider.baseURL ?? 'http://localhost:1234/v1'))
+  const apiRoot = normalizedBaseURL.endsWith('/v1') ? normalizedBaseURL.slice(0, -3) : normalizedBaseURL
+  const { systemPrompt, input } = lmStudioNativeContent(messages)
+  const reasoningEffort = options.reasoningEffort ?? provider.reasoningEffort
+  const reasoning = reasoningEffort === 'none'
+    ? 'off'
+    : ['low', 'medium', 'high'].includes(reasoningEffort)
+      ? reasoningEffort
+      : reasoningEffort === 'max'
+        ? 'high'
+        : undefined
+  const body = {
+    model: provider.model,
+    input,
+    store: false,
+    max_output_tokens: options.maxTokens ?? provider.maxTokens ?? 4096,
+    ...(systemPrompt === '' ? {} : { system_prompt: systemPrompt }),
+    ...(reasoning === undefined ? {} : { reasoning }),
+    ...(typeof options.temperature === 'number' ? { temperature: options.temperature } : {}),
+    ...(typeof options.top_p === 'number' ? { top_p: options.top_p } : {}),
+  }
+  const response = await fetch(`${apiRoot}/api/v1/chat`, {
+    method: 'POST',
+    headers: { 'content-type': 'application/json' },
+    body: JSON.stringify(body),
+    ...(options.signal === undefined ? {} : { signal: options.signal }),
+  })
+  if (!response.ok) {
+    const detail = (await readResponseTextBounded(
+      response,
+      ERROR_RESPONSE_MAX_BYTES,
+      { label: 'LM Studio native error response' },
+    ).catch(() => '')).slice(0, 300)
+    const error = new Error(`http provider "${provider.name}": ${response.status} ${detail}`)
+    error.status = response.status
+    error.code = kindForHttpStatus(response.status) ?? 'HTTP_PROVIDER_FAILED'
+    throw error
+  }
+  const data = await readResponseJsonBounded(
+    response,
+    MODEL_RESPONSE_MAX_BYTES,
+    { label: 'LM Studio native response' },
+  )
+  const text = Array.isArray(data?.output)
+    ? data.output
+        .filter((item) => item?.type === 'message' && typeof item.content === 'string')
+        .map((item) => item.content)
+        .join('\n')
+        .trim()
+    : ''
+  if (text === '' && Number(data?.stats?.total_output_tokens) >= body.max_output_tokens) {
+    const error = new Error(`http provider "${provider.name}": empty answer after exhausting the completion budget`)
+    error.code = 'VISION_EMPTY_RESPONSE'
+    throw error
+  }
+  return text
+}
+
 export async function callLocalBackend(provider, messages, options = {}) {
   const maxTokens = options.maxTokens ?? provider.maxTokens ?? 2048
   const sampling = {
     ...(typeof provider.temperature === 'number' ? { temperature: provider.temperature } : {}),
     ...(typeof provider.top_p === 'number' ? { top_p: provider.top_p } : {}),
   }
+  if (provider.format === 'lmstudio') {
+    return callLmStudioNative(provider, messages, {
+      maxTokens,
+      signal: options.signal,
+      ...(typeof provider.reasoningEffort === 'string' ? { reasoningEffort: provider.reasoningEffort } : {}),
+      ...sampling,
+    })
+  }
   if (provider.format === 'anthropic') {
     const system = []
     const wire = []
@@ -2182,11 +2287,16 @@ export async function callOpenAICompatible(provider, messages, options = {}) {
     MODEL_RESPONSE_MAX_BYTES,
     { label: `http provider \"${provider.name}\" response` },
   )
-  const content = data && data.choices && data.choices[0] && data.choices[0].message
-    ? data.choices[0].message.content
-    : undefined
+  const choice = data && data.choices && data.choices[0] ? data.choices[0] : undefined
+  const content = choice && choice.message ? choice.message.content : undefined
   if (typeof content !== 'string') throw new Error(`http provider "${provider.name}": unexpected response shape`)
-  return content.trim()
+  const text = content.trim()
+  if (text === '' && choice?.finish_reason === 'length' && (provider.name === 'local-ollama' || provider.name === 'local-lmstudio')) {
+    const error = new Error(`http provider "${provider.name}": empty answer after exhausting the completion budget`)
+    error.code = 'VISION_EMPTY_RESPONSE'
+    throw error
+  }
+  return text
 }
 
 /**

package/lib/dsh-host-capabilities.js

@@ -43,6 +43,7 @@ export function inspectDshHostCapabilities(ctx) {
   const llm = serviceOf(ctx, 'llm')
   const jobs = serviceOf(ctx, 'jobs')
   const settings = serviceOf(ctx, 'settings')
+  const configEditor = serviceOf(ctx, 'configEditor')
   const tools = serviceOf(ctx, 'tools')
 
   let maxImageDimension = false
@@ -57,6 +58,10 @@ export function inspectDshHostCapabilities(ctx) {
   const prepareCall = hasFunction(llm, 'prepareCall')
   const toolRegistration = hasFunction(tools, 'register') || hasFunction(tools, 'registerTool')
   const toolExecution = hasFunction(tools, 'call') || hasFunction(tools, 'execute') || hasFunction(tools, 'invoke')
+  const settingsLiveNamespace = hasFunction(settings, 'register')
+    || (hasFunction(settings, 'describe')
+      && hasFunction(configEditor, 'configuration')
+      && hasFunction(configEditor, 'edit'))
 
   return Object.freeze({
     batchAttachments: hasFunction(attachments, 'saveImages'),
@@ -70,10 +75,11 @@ export function inspectDshHostCapabilities(ctx) {
     // No stable read-only client-surface capability flag exists across the
     // supported Host window yet.
     surfaceReplacement: UNKNOWN,
-    // DSH SettingsProvider.register() is the public live-namespace seam; get()
-    // and watch() belong to the returned SettingsScope rather than the service.
-    // Detect the stable service method and do not manufacture a registration.
-    settingsLiveNamespace: hasFunction(settings, 'register'),
+    // Older Hosts expose SettingsProvider.register() directly. DSH 0.1.7's
+    // reviewed replacement is SettingsForms.describe() plus ConfigEditor
+    // configuration()/edit(), which DVR adapts into the same live namespace.
+    // Probe only readable method presence; never manufacture a registration or write.
+    settingsLiveNamespace,
     // Browser exposure is a presentation capability and is not safely implied
     // by the Host settings service alone.
     settingsWebExposure: UNKNOWN,