dsh-chief 1.4.0 → 1.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/chief-models.js CHANGED
@@ -10,6 +10,8 @@ import {
10
10
  SUMMARY_MAX_TOKENS,
11
11
  SUMMARY_RETRY_MAX_TOKENS,
12
12
  clamp,
13
+ normalizeThinkingEffort,
14
+ resolveAlignedModelEffort,
13
15
  resolveWorkspaceConfig,
14
16
  loadChiefState,
15
17
  mutateChiefState,
@@ -180,6 +182,75 @@ function profileDefaultModelRef(runtime) {
180
182
  return ''
181
183
  }
182
184
 
185
+ const modelEffortCache = new Map()
186
+
187
+ export function decideAdaptiveChiefEffort({
188
+ reasonKind = '',
189
+ userText = '',
190
+ isBlankContext = false,
191
+ tactical = {},
192
+ } = {}) {
193
+ const rk = String(reasonKind || '').trim().toLowerCase()
194
+ const txt = String(userText || '').trim()
195
+
196
+ if (rk === 'chat_summary' || rk === 'dream_roam' || rk === 'quick_status' || rk === 'auto_summary') {
197
+ return 'off'
198
+ }
199
+ if (rk === 'hazard_diagnosis' || rk === 'multi_wall_collision') {
200
+ return 'xhigh'
201
+ }
202
+ if (
203
+ rk === 'blank_explore'
204
+ || rk === 'wall_blocked'
205
+ || rk === 'strong_intervention'
206
+ || rk === 'precipitate'
207
+ || rk === 'distill'
208
+ || rk === 'craft_wall'
209
+ || isBlankContext === true
210
+ ) {
211
+ if (tactical && tactical.intervention && /(?:架构|根因|死锁|并发|重构|安全审计)/.test(txt)) {
212
+ return 'xhigh'
213
+ }
214
+ return 'high'
215
+ }
216
+ if (txt && /(?:深度|全面|彻底|根因|重构|架构|死锁|并发|安全审计|复盘|仔细想|好好想|正则墙|物理墙)/.test(txt)) {
217
+ return 'high'
218
+ }
219
+ if (rk === 'ask_chat' && txt && txt.length <= 28 && /^(?:在吗|进度|到哪一步|怎么样|状态|你好|看看)/.test(txt)) {
220
+ return 'low'
221
+ }
222
+ return 'medium'
223
+ }
224
+
225
+ export async function resolveRouteCapabilities(llm, provider, model) {
226
+ const ref = provider && model ? `${provider}/${model}` : ''
227
+ if (ref && modelEffortCache.has(ref)) {
228
+ const cached = modelEffortCache.get(ref)
229
+ if (Date.now() - cached.at < CATALOG_TTL_MS * 6) return cached.data
230
+ }
231
+ const data = { defaultMaxTokens: 0, efforts: [], defaultEffort: '' }
232
+ try {
233
+ if (llm && typeof llm.resolveModelInfo === 'function' && provider && model) {
234
+ const info = await llm.resolveModelInfo(provider, model)
235
+ if (typeof info?.defaultMaxTokens === 'number' && info.defaultMaxTokens > 0) {
236
+ data.defaultMaxTokens = info.defaultMaxTokens
237
+ }
238
+ if (Array.isArray(info?.reasoning?.efforts)) {
239
+ data.efforts = info.reasoning.efforts.map(e => (
240
+ typeof e === 'string'
241
+ ? { id: e, name: e }
242
+ : { id: String(e?.id || ''), name: String(e?.name || e?.id || ''), description: String(e?.description || '') }
243
+ )).filter(e => e.id)
244
+ }
245
+ if (typeof info?.reasoning?.defaultEffort === 'string') {
246
+ data.defaultEffort = info.reasoning.defaultEffort
247
+ }
248
+ }
249
+ } catch {}
250
+ if (ref) modelEffortCache.set(ref, { at: Date.now(), data })
251
+ return data
252
+ }
253
+
183
254
  export async function listChiefModels(deps, fresh = false, sessionId = '') {
184
255
  const llm = deps.getLlm ? deps.getLlm() : null
185
256
  const runtime = deps.runtime
@@ -297,6 +368,16 @@ export async function listChiefModels(deps, fresh = false, sessionId = '') {
297
368
 
298
369
  const picked = pickRouteFromItems(items, selectedRef, followRef, deps, sessionId, registeredProviderIds)
299
370
  const resolvedRef = (picked && picked.ref) || selectedRef || followRef || ''
371
+ let supportedEfforts = []
372
+ let defaultEffort = ''
373
+ if (resolvedRef) {
374
+ const parsedResolved = splitModelRef(resolvedRef)
375
+ if (parsedResolved) {
376
+ const caps = await resolveRouteCapabilities(llm, parsedResolved.provider, parsedResolved.model)
377
+ supportedEfforts = caps.efforts || []
378
+ defaultEffort = caps.defaultEffort || ''
379
+ }
380
+ }
300
381
  return {
301
382
  available: items.length > 0 || Boolean(resolvedRef),
302
383
  selectedRef,
@@ -305,6 +386,8 @@ export async function listChiefModels(deps, fresh = false, sessionId = '') {
305
386
  defaultRef: followRef,
306
387
  resolvedRef,
307
388
  resolvedLabel: formatChiefModelLabel(resolvedRef, items, rawProviders, true),
389
+ supportedEfforts,
390
+ defaultEffort,
308
391
  models: items,
309
392
  }
310
393
  }
@@ -397,21 +480,11 @@ function pickRouteFromItems(items, selectedRef, followRef, deps, sessionId, regi
397
480
  return queue[0] || null
398
481
  }
399
482
 
400
- async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MAX_TOKENS, sessionId = '', imageBlocks = []) {
401
- let effort
402
- let modelDefaultMaxTokens
403
- try {
404
- if (typeof llm.resolveModelInfo === 'function') {
405
- const info = await llm.resolveModelInfo(route.provider, route.model)
406
- if (typeof info?.defaultMaxTokens === 'number' && info.defaultMaxTokens > 0) {
407
- modelDefaultMaxTokens = info.defaultMaxTokens
408
- }
409
- const efforts = info?.reasoning?.efforts
410
- if (Array.isArray(efforts) && efforts.length > 0) {
411
- effort = efforts.some(e => e.id === 'off') ? 'off' : efforts[0]?.id
412
- }
413
- }
414
- } catch {}
483
+ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MAX_TOKENS, sessionId = '', imageBlocks = [], desiredEffort = 'medium') {
484
+ const caps = await resolveRouteCapabilities(llm, route.provider, route.model)
485
+ const modelDefaultMaxTokens = caps.defaultMaxTokens || 0
486
+ const efforts = Array.isArray(caps.efforts) ? caps.efforts : []
487
+ const effort = resolveAlignedModelEffort(desiredEffort, efforts)
415
488
 
416
489
  const targetBudget = Math.max(Number(maxTokens) || 0, SUMMARY_MAX_TOKENS)
417
490
  const firstBudget = modelDefaultMaxTokens ? Math.min(modelDefaultMaxTokens, targetBudget) : targetBudget
@@ -450,7 +523,7 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
450
523
  .map(b => b.text)
451
524
  .join('\n')
452
525
  .trim()
453
- return { text, finishKind: finish?.kind || '', modelRef: route.ref }
526
+ return { text, finishKind: finish?.kind || '', modelRef: route.ref, thinkingEffort: effort || 'none' }
454
527
  }
455
528
 
456
529
  // Pure standalone fallback stream reader (for unit tests / lightweight hosts)
@@ -472,18 +545,18 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
472
545
  }
473
546
  }
474
547
  }
475
- return { text: parts.join('').trim(), finishKind, modelRef: route.ref }
548
+ return { text: parts.join('').trim(), finishKind, modelRef: route.ref, thinkingEffort: effort || 'none' }
476
549
  }
477
550
 
478
- let lastRes = { text: '', finishKind: '', modelRef: route.ref }
551
+ let lastRes = { text: '', finishKind: '', modelRef: route.ref, thinkingEffort: effort || 'none' }
479
552
  for (let i = 0; i < budgets.length; i += 1) {
480
553
  const res = await runOnce(budgets[i])
481
554
  lastRes = res
482
555
  if (res.text && res.finishKind !== 'max-tokens') {
483
- return { text: res.text, modelRef: res.modelRef }
556
+ return { text: res.text, modelRef: res.modelRef, thinkingEffort: res.thinkingEffort }
484
557
  }
485
558
  }
486
- return { text: lastRes.text, modelRef: lastRes.modelRef }
559
+ return { text: lastRes.text, modelRef: lastRes.modelRef, thinkingEffort: lastRes.thinkingEffort }
487
560
  }
488
561
 
489
562
  async function filterVisionRoutes(llm, queue) {
@@ -515,14 +588,51 @@ export async function callChiefLlm(deps, {
515
588
  maxTokens = SUMMARY_MAX_TOKENS,
516
589
  imageBlocks = [],
517
590
  allowTextFallbackOnVisionMiss = false,
591
+ reasonKind = '',
592
+ userText = '',
593
+ isBlankContext = false,
594
+ tactical = null,
595
+ effortOverride = '',
518
596
  }) {
519
597
  const safeImages = Array.isArray(imageBlocks) ? imageBlocks : []
598
+ const rootSid = (deps && deps.resolveRootSessionId ? deps.resolveRootSessionId(sessionId) : '') || String(sessionId || '').trim()
599
+ const st = loadChiefState()
600
+ const effCfg = resolveWorkspaceConfig(st, rootSid)
601
+ const cfgEffort = normalizeThinkingEffort(effortOverride || effCfg.thinkingEffort || 'auto')
602
+ const thinkingEffortMode = cfgEffort === 'auto' ? 'auto' : 'manual'
603
+ const desiredEffort = cfgEffort === 'auto'
604
+ ? decideAdaptiveChiefEffort({
605
+ reasonKind,
606
+ userText: userText || prompt,
607
+ isBlankContext,
608
+ tactical: tactical || effCfg.tacticalModes || {},
609
+ })
610
+ : cfgEffort
611
+
520
612
  if (deps && typeof deps.callLlmOverride === 'function') {
521
- return deps.callLlmOverride({ sessionId, system, prompt, maxTokens, imageBlocks: safeImages, allowTextFallbackOnVisionMiss })
613
+ const overrideRes = await deps.callLlmOverride({
614
+ sessionId,
615
+ system,
616
+ prompt,
617
+ maxTokens,
618
+ imageBlocks: safeImages,
619
+ allowTextFallbackOnVisionMiss,
620
+ reasonKind,
621
+ desiredEffort,
622
+ thinkingEffortMode,
623
+ })
624
+ if (overrideRes && typeof overrideRes === 'object') {
625
+ return {
626
+ thinkingEffort: overrideRes.thinkingEffort || desiredEffort,
627
+ thinkingEffortMode: overrideRes.thinkingEffortMode || thinkingEffortMode,
628
+ ...overrideRes,
629
+ }
630
+ }
631
+ return overrideRes
522
632
  }
523
633
  const llm = deps.getLlm ? deps.getLlm() : null
524
634
  if (!llm || typeof llm.stream !== 'function') {
525
- return { ok: false, error: '当前环境未挂载 LLM 服务' }
635
+ return { ok: false, error: '当前环境未挂载 LLM 服务', thinkingEffort: desiredEffort, thinkingEffortMode }
526
636
  }
527
637
  const catalog = await listChiefModels(deps, false, sessionId)
528
638
  const regSet = typeof llm.listProviders === 'function'
@@ -530,7 +640,7 @@ export async function callChiefLlm(deps, {
530
640
  : null
531
641
  const baseQueue = buildRouteQueue(catalog.models, catalog.selectedRef, catalog.followRef || catalog.defaultRef, deps, sessionId, regSet)
532
642
  if (!baseQueue.length) {
533
- return { ok: false, error: '未找到可用的参谋长模型,请在 DSH 设置中配置模型供应商' }
643
+ return { ok: false, error: '未找到可用的参谋长模型,请在 DSH 设置中配置模型供应商', thinkingEffort: desiredEffort, thinkingEffortMode }
534
644
  }
535
645
 
536
646
  let queue = baseQueue
@@ -548,6 +658,8 @@ export async function callChiefLlm(deps, {
548
658
  return {
549
659
  ok: false,
550
660
  error: '当前参谋长及备选模型均不支持图片输入(仅纯文本),请在右上角切换支持视觉的多模态模型(如 Gemini / Claude / GPT-4o / Qwen-VL 等)后再试。',
661
+ thinkingEffort: desiredEffort,
662
+ thinkingEffortMode,
551
663
  }
552
664
  }
553
665
  }
@@ -555,22 +667,40 @@ export async function callChiefLlm(deps, {
555
667
  let lastError = '模型返回内容为空'
556
668
  for (const route of queue.slice(0, 4)) {
557
669
  try {
558
- const res = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, sendImages)
559
- if (res.text) return { ok: true, text: res.text, modelRef: res.modelRef, degradedToTextOnly }
670
+ const res = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, sendImages, desiredEffort)
671
+ if (res.text) {
672
+ return {
673
+ ok: true,
674
+ text: res.text,
675
+ modelRef: res.modelRef,
676
+ degradedToTextOnly,
677
+ thinkingEffort: res.thinkingEffort || desiredEffort,
678
+ thinkingEffortMode,
679
+ }
680
+ }
560
681
  lastError = `${route.ref} 返回内容为空`
561
682
  markRouteFailure(route, new Error(lastError))
562
683
  } catch (err) {
563
684
  lastError = String(err && err.message ? err.message : err)
564
685
  if (sendImages.length > 0 && allowTextFallbackOnVisionMiss) {
565
686
  try {
566
- const textRes = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, [])
567
- if (textRes.text) return { ok: true, text: textRes.text, modelRef: textRes.modelRef, degradedToTextOnly: true }
687
+ const textRes = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, [], desiredEffort)
688
+ if (textRes.text) {
689
+ return {
690
+ ok: true,
691
+ text: textRes.text,
692
+ modelRef: textRes.modelRef,
693
+ degradedToTextOnly: true,
694
+ thinkingEffort: textRes.thinkingEffort || desiredEffort,
695
+ thinkingEffortMode,
696
+ }
697
+ }
568
698
  } catch {}
569
699
  }
570
700
  markRouteFailure(route, err)
571
701
  }
572
702
  }
573
- return { ok: false, error: lastError, degradedToTextOnly }
703
+ return { ok: false, error: lastError, degradedToTextOnly, thinkingEffort: desiredEffort, thinkingEffortMode }
574
704
  }
575
705
 
576
706
  /* -------------------------------------------------------------------------- */