dsh-chief 1.5.0 → 1.7.0

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.
@@ -546,6 +546,7 @@ export async function recommendAndSetCompactThreshold(deps, { sessionId = '' } =
546
546
 
547
547
  const llmRes = await callChiefLlm(deps, {
548
548
  sessionId: rootId,
549
+ reasonKind: 'recommend_compact',
549
550
  system: COMPACT_ADVISOR_SYSTEM,
550
551
  prompt: advisorPrompt,
551
552
  maxTokens: 320,
package/chief-executor.js CHANGED
@@ -521,6 +521,7 @@ export async function executeChiefToolCall(deps, call, execCtx = {}) {
521
521
  const rootId = String(execCtx.rootId || '')
522
522
  const board = execCtx.board || {}
523
523
  const readOnlyToolsOnly = Boolean(execCtx.readOnlyToolsOnly)
524
+ const safetyGuardEnabled = execCtx.safetyGuardEnabled !== false
524
525
  const approvedOnce = Boolean(execCtx.approvedOnce)
525
526
  const capabilityScope = execCtx.capabilityScope || null
526
527
 
@@ -563,7 +564,7 @@ export async function executeChiefToolCall(deps, call, execCtx = {}) {
563
564
  const resolvedAbsTarget = rawTargetPath ? path.resolve(cwd, rawTargetPath) : ''
564
565
  const rawCommand = tool === 'run_command' ? String(args.command || args.cmd || '').trim() : ''
565
566
 
566
- // Stage 3: 安全卫兵硬性规则(优先于任何权限等级、策略插件与逐项批准)
567
+ // Stage 3: 安全卫兵硬性规则(优先于任何权限等级、策略插件与逐项批准;受人类可控 safetyGuardEnabled 开关控制)
567
568
  const guardRes = evaluateSafetyGuardHardRules({
568
569
  tool,
569
570
  opClass,
@@ -572,6 +573,7 @@ export async function executeChiefToolCall(deps, call, execCtx = {}) {
572
573
  args,
573
574
  cwd,
574
575
  board,
576
+ safetyGuardEnabled,
575
577
  })
576
578
  if (guardRes.blocked) {
577
579
  writeAudit({
@@ -1093,6 +1095,7 @@ export async function runChiefAgenticLoop(deps, {
1093
1095
  board = null,
1094
1096
  maxSteps = 4,
1095
1097
  readOnlyToolsOnly = false,
1098
+ safetyGuardEnabled = true,
1096
1099
  disableTools = false,
1097
1100
  lazyDeferWorkspaceWritesToLead = false,
1098
1101
  callLlmFn,
@@ -1182,6 +1185,7 @@ export async function runChiefAgenticLoop(deps, {
1182
1185
  args: (c && c.args && typeof c.args === 'object') ? c.args : {},
1183
1186
  cwd,
1184
1187
  board,
1188
+ safetyGuardEnabled,
1185
1189
  })
1186
1190
  if (!preDef || preDef.decision !== 'deny') {
1187
1191
  if (typeof recordAuditFn === 'function') {
@@ -1222,6 +1226,7 @@ export async function runChiefAgenticLoop(deps, {
1222
1226
  policyPlugins,
1223
1227
  board,
1224
1228
  readOnlyToolsOnly,
1229
+ safetyGuardEnabled,
1225
1230
  upsertArchiveFn,
1226
1231
  compactNowFn,
1227
1232
  queueApprovalFn,
package/chief-models.js CHANGED
@@ -15,6 +15,8 @@ import {
15
15
  resolveWorkspaceConfig,
16
16
  loadChiefState,
17
17
  mutateChiefState,
18
+ checkManagedBudgetBeforeCall,
19
+ recordChiefLlmUsage,
18
20
  } from './chief-state.js'
19
21
 
20
22
  let dshLlmModulePromise = null
@@ -480,6 +482,36 @@ function pickRouteFromItems(items, selectedRef, followRef, deps, sessionId, regi
480
482
  return queue[0] || null
481
483
  }
482
484
 
485
+ export function estimateTokensRough(text = '') {
486
+ const s = String(text || '')
487
+ if (!s) return 0
488
+ let cjk = 0
489
+ for (let i = 0; i < s.length; i += 1) {
490
+ const code = s.charCodeAt(i)
491
+ if (code >= 0x4E00 && code <= 0x9FFF) cjk += 1
492
+ }
493
+ const rest = s.length - cjk
494
+ return Math.max(1, Math.round(cjk * 0.72 + rest * 0.28))
495
+ }
496
+
497
+ function extractUsageMetrics(rawUsage, promptText = '', outputText = '') {
498
+ const u = rawUsage && typeof rawUsage === 'object' ? rawUsage : null
499
+ const inRaw = u ? Number(u.inputTokens ?? u.promptTokens ?? u.prompt_tokens ?? u.input_tokens ?? 0) : 0
500
+ const outRaw = u ? Number(u.outputTokens ?? u.completionTokens ?? u.completion_tokens ?? u.output_tokens ?? 0) : 0
501
+ if (Number.isFinite(inRaw) && Number.isFinite(outRaw) && (inRaw > 0 || outRaw > 0)) {
502
+ return {
503
+ inputTokens: Math.max(0, Math.round(inRaw)),
504
+ outputTokens: Math.max(0, Math.round(outRaw)),
505
+ estimated: false,
506
+ }
507
+ }
508
+ return {
509
+ inputTokens: estimateTokensRough(promptText),
510
+ outputTokens: estimateTokensRough(outputText),
511
+ estimated: true,
512
+ }
513
+ }
514
+
483
515
  async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MAX_TOKENS, sessionId = '', imageBlocks = [], desiredEffort = 'medium') {
484
516
  const caps = await resolveRouteCapabilities(llm, route.provider, route.model)
485
517
  const modelDefaultMaxTokens = caps.defaultMaxTokens || 0
@@ -491,6 +523,7 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
491
523
  const retryBudget = firstBudget < targetBudget ? targetBudget : SUMMARY_RETRY_MAX_TOKENS
492
524
  const budgets = retryBudget > firstBudget ? [firstBudget, retryBudget] : [firstBudget]
493
525
  const safeImages = Array.isArray(imageBlocks) ? imageBlocks : []
526
+ const promptBasis = `${system || ''}\n${userText || ''}`
494
527
 
495
528
  const runOnce = async (budget) => {
496
529
  const userContent = [...safeImages, { type: 'text', text: userText || '请查看附图并回答。' }]
@@ -508,8 +541,12 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
508
541
  const dshLlm = await loadDshLlm()
509
542
  if (dshLlm && typeof dshLlm.BlockAssembler === 'function') {
510
543
  const assembler = new dshLlm.BlockAssembler()
544
+ let streamUsage = null
511
545
  for await (const chunk of llm.stream(request)) {
512
546
  assembler.push(chunk)
547
+ if (chunk && typeof chunk === 'object' && (chunk.usage || chunk.type === 'usage')) {
548
+ streamUsage = chunk.usage || chunk
549
+ }
513
550
  }
514
551
  const finish = assembler.finish
515
552
  if (finish && (finish.kind === 'error' || finish.kind === 'aborted')) {
@@ -523,14 +560,19 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
523
560
  .map(b => b.text)
524
561
  .join('\n')
525
562
  .trim()
526
- return { text, finishKind: finish?.kind || '', modelRef: route.ref, thinkingEffort: effort || 'none' }
563
+ const usage = extractUsageMetrics(assembler.usage || finish?.usage || finish?.reason?.usage || streamUsage, promptBasis, text)
564
+ return { text, finishKind: finish?.kind || '', modelRef: route.ref, thinkingEffort: effort || 'none', usage }
527
565
  }
528
566
 
529
567
  // Pure standalone fallback stream reader (for unit tests / lightweight hosts)
530
568
  const parts = []
531
569
  let finishKind = ''
570
+ let streamUsage = null
532
571
  for await (const chunk of llm.stream(request)) {
533
572
  if (!chunk || typeof chunk !== 'object') continue
573
+ if (chunk.usage || chunk.type === 'usage') {
574
+ streamUsage = chunk.usage || chunk
575
+ }
534
576
  if (chunk.type === 'text-delta' || chunk.type === 'text_delta') {
535
577
  if (typeof chunk.text === 'string') parts.push(chunk.text)
536
578
  else if (typeof chunk.delta === 'string') parts.push(chunk.delta)
@@ -538,6 +580,7 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
538
580
  else if (typeof chunk.content === 'string') parts.push(chunk.content)
539
581
  else if (chunk.type === 'finish') {
540
582
  finishKind = chunk.reason?.kind || chunk.finishReason || ''
583
+ if (chunk.reason?.usage) streamUsage = chunk.reason.usage
541
584
  if (finishKind === 'error') {
542
585
  const err = new Error(chunk.reason?.failure?.message || 'LLM stream error')
543
586
  err.failure = chunk.reason?.failure
@@ -545,18 +588,26 @@ async function callOneRoute(llm, route, system, userText, maxTokens = SUMMARY_MA
545
588
  }
546
589
  }
547
590
  }
548
- return { text: parts.join('').trim(), finishKind, modelRef: route.ref, thinkingEffort: effort || 'none' }
591
+ const text = parts.join('').trim()
592
+ const usage = extractUsageMetrics(streamUsage, promptBasis, text)
593
+ return { text, finishKind, modelRef: route.ref, thinkingEffort: effort || 'none', usage }
549
594
  }
550
595
 
551
- let lastRes = { text: '', finishKind: '', modelRef: route.ref, thinkingEffort: effort || 'none' }
596
+ let lastRes = {
597
+ text: '',
598
+ finishKind: '',
599
+ modelRef: route.ref,
600
+ thinkingEffort: effort || 'none',
601
+ usage: extractUsageMetrics(null, promptBasis, ''),
602
+ }
552
603
  for (let i = 0; i < budgets.length; i += 1) {
553
604
  const res = await runOnce(budgets[i])
554
605
  lastRes = res
555
606
  if (res.text && res.finishKind !== 'max-tokens') {
556
- return { text: res.text, modelRef: res.modelRef, thinkingEffort: res.thinkingEffort }
607
+ return { text: res.text, modelRef: res.modelRef, thinkingEffort: res.thinkingEffort, usage: res.usage }
557
608
  }
558
609
  }
559
- return { text: lastRes.text, modelRef: lastRes.modelRef, thinkingEffort: lastRes.thinkingEffort }
610
+ return { text: lastRes.text, modelRef: lastRes.modelRef, thinkingEffort: lastRes.thinkingEffort, usage: lastRes.usage }
560
611
  }
561
612
 
562
613
  async function filterVisionRoutes(llm, queue) {
@@ -609,6 +660,20 @@ export async function callChiefLlm(deps, {
609
660
  })
610
661
  : cfgEffort
611
662
 
663
+ const budgetCheck = checkManagedBudgetBeforeCall(rootSid, reasonKind)
664
+ if (budgetCheck && !budgetCheck.allowed) {
665
+ return {
666
+ ok: false,
667
+ budgetExceeded: true,
668
+ budgetTripped: true,
669
+ error: budgetCheck.reason || '已触发托管硬预算熔断',
670
+ thinkingEffort: desiredEffort,
671
+ thinkingEffortMode,
672
+ }
673
+ }
674
+
675
+ const promptBasis = `${system || ''}\n${prompt || ''}`
676
+
612
677
  if (deps && typeof deps.callLlmOverride === 'function') {
613
678
  const overrideRes = await deps.callLlmOverride({
614
679
  sessionId,
@@ -622,9 +687,20 @@ export async function callChiefLlm(deps, {
622
687
  thinkingEffortMode,
623
688
  })
624
689
  if (overrideRes && typeof overrideRes === 'object') {
690
+ const usage = extractUsageMetrics(overrideRes.usage, promptBasis, overrideRes.text || '')
691
+ const ledgerRec = recordChiefLlmUsage(rootSid, {
692
+ reasonKind,
693
+ modelRef: overrideRes.modelRef || 'override',
694
+ inputTokens: usage.inputTokens,
695
+ outputTokens: usage.outputTokens,
696
+ estimated: usage.estimated,
697
+ ok: overrideRes.ok !== false,
698
+ })
625
699
  return {
626
700
  thinkingEffort: overrideRes.thinkingEffort || desiredEffort,
627
701
  thinkingEffortMode: overrideRes.thinkingEffortMode || thinkingEffortMode,
702
+ usage,
703
+ ...(ledgerRec && ledgerRec.tripped ? { budgetTripped: true } : {}),
628
704
  ...overrideRes,
629
705
  }
630
706
  }
@@ -669,6 +745,15 @@ export async function callChiefLlm(deps, {
669
745
  try {
670
746
  const res = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, sendImages, desiredEffort)
671
747
  if (res.text) {
748
+ const usage = res.usage || extractUsageMetrics(null, promptBasis, res.text)
749
+ const ledgerRec = recordChiefLlmUsage(rootSid, {
750
+ reasonKind,
751
+ modelRef: res.modelRef,
752
+ inputTokens: usage.inputTokens,
753
+ outputTokens: usage.outputTokens,
754
+ estimated: usage.estimated,
755
+ ok: true,
756
+ })
672
757
  return {
673
758
  ok: true,
674
759
  text: res.text,
@@ -676,6 +761,8 @@ export async function callChiefLlm(deps, {
676
761
  degradedToTextOnly,
677
762
  thinkingEffort: res.thinkingEffort || desiredEffort,
678
763
  thinkingEffortMode,
764
+ usage,
765
+ ...(ledgerRec && ledgerRec.tripped ? { budgetTripped: true } : {}),
679
766
  }
680
767
  }
681
768
  lastError = `${route.ref} 返回内容为空`
@@ -686,6 +773,15 @@ export async function callChiefLlm(deps, {
686
773
  try {
687
774
  const textRes = await callOneRoute(llm, route, system, prompt, maxTokens, sessionId, [], desiredEffort)
688
775
  if (textRes.text) {
776
+ const usage = textRes.usage || extractUsageMetrics(null, promptBasis, textRes.text)
777
+ const ledgerRec = recordChiefLlmUsage(rootSid, {
778
+ reasonKind,
779
+ modelRef: textRes.modelRef,
780
+ inputTokens: usage.inputTokens,
781
+ outputTokens: usage.outputTokens,
782
+ estimated: usage.estimated,
783
+ ok: true,
784
+ })
689
785
  return {
690
786
  ok: true,
691
787
  text: textRes.text,
@@ -693,6 +789,8 @@ export async function callChiefLlm(deps, {
693
789
  degradedToTextOnly: true,
694
790
  thinkingEffort: textRes.thinkingEffort || desiredEffort,
695
791
  thinkingEffortMode,
792
+ usage,
793
+ ...(ledgerRec && ledgerRec.tripped ? { budgetTripped: true } : {}),
696
794
  }
697
795
  }
698
796
  } catch {}
@@ -700,6 +798,15 @@ export async function callChiefLlm(deps, {
700
798
  markRouteFailure(route, err)
701
799
  }
702
800
  }
801
+ const failUsage = extractUsageMetrics(null, promptBasis, '')
802
+ recordChiefLlmUsage(rootSid, {
803
+ reasonKind,
804
+ modelRef: queue[0] ? queue[0].ref : '',
805
+ inputTokens: failUsage.inputTokens,
806
+ outputTokens: 0,
807
+ estimated: true,
808
+ ok: false,
809
+ })
703
810
  return { ok: false, error: lastError, degradedToTextOnly, thinkingEffort: desiredEffort, thinkingEffortMode }
704
811
  }
705
812