dsh-chief 1.2.1 → 1.3.0

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.
@@ -0,0 +1,578 @@
1
+ /**
2
+ * chief-models.js
3
+ *
4
+ * Model catalog discovery, route cooldown tracking, and ctx.llm invocation for DSH Chief.
5
+ */
6
+
7
+ import {
8
+ MODEL_TIMEOUT_MS,
9
+ CATALOG_TTL_MS,
10
+ SUMMARY_MAX_TOKENS,
11
+ SUMMARY_RETRY_MAX_TOKENS,
12
+ clamp,
13
+ resolveWorkspaceConfig,
14
+ loadChiefState,
15
+ mutateChiefState,
16
+ } from './chief-state.js'
17
+
18
+ let dshLlmModulePromise = null
19
+
20
+ export async function loadDshLlm() {
21
+ if (!dshLlmModulePromise) {
22
+ dshLlmModulePromise = import('@deepseek-ai/dsh-llm').catch(() => null)
23
+ }
24
+ return dshLlmModulePromise
25
+ }
26
+
27
+ let modelCatalogCache = { at: 0, items: [] }
28
+
29
+ export function invalidateChiefModelCatalog() {
30
+ modelCatalogCache = { at: 0, items: [] }
31
+ unusableRouteUntil.clear()
32
+ }
33
+
34
+ function splitModelRef(ref) {
35
+ const clean = String(ref || '').trim()
36
+ const idx = clean.indexOf('/')
37
+ if (idx <= 0 || idx >= clean.length - 1) return null
38
+ return { provider: clean.slice(0, idx), model: clean.slice(idx + 1), ref: clean }
39
+ }
40
+
41
+ function stripVendorPrefixFromModelId(modelId) {
42
+ const raw = String(modelId || '').trim()
43
+ if (!raw.includes('/')) return raw
44
+ const parts = raw.split('/').filter(Boolean)
45
+ return parts.length > 1 ? parts.slice(1).join('/') : raw
46
+ }
47
+
48
+ function shortProviderTag(providerId, providerName) {
49
+ const pid = String(providerId || '').toLowerCase()
50
+ const pname = String(providerName || '').trim()
51
+ if (pid === 'commandcode' || /command\s*code/i.test(pname)) return 'Command Code'
52
+ if (pid.includes('account') || /account/i.test(pname)) return pname || 'DeepSeek Account'
53
+ if (pid === 'our-free-model') return 'Free Model'
54
+ return pname || providerId || ''
55
+ }
56
+
57
+ export function formatChiefModelLabel(ref, items = [], providers = [], compact = false) {
58
+ const parsed = splitModelRef(ref)
59
+ if (!parsed) return compact ? '自动跟随' : ''
60
+ const hit = Array.isArray(items) ? items.find(it => it && it.ref === parsed.ref) : null
61
+ const provObj = Array.isArray(providers) ? providers.find(p => p && p.id === parsed.provider) : null
62
+ const provLabel = shortProviderTag(parsed.provider, (hit && hit.providerName) || (provObj && (provObj.name || provObj.label)) || parsed.provider)
63
+ const rawModelName = (hit && hit.name) ? String(hit.name).trim() : stripVendorPrefixFromModelId(parsed.model)
64
+
65
+ if (compact) {
66
+ // Keep channel crystal-clear so Command Code DS is never confused with official DeepSeek API
67
+ const shortProv = parsed.provider === 'commandcode' ? 'CC' : provLabel
68
+ const alreadyTagged = /\(CC\)/i.test(rawModelName) && shortProv === 'CC'
69
+ const base = alreadyTagged ? rawModelName : (shortProv ? `${shortProv} · ${rawModelName}` : rawModelName)
70
+ return clamp(base, 24)
71
+ }
72
+ const alreadyHasProv = provLabel && rawModelName.toLowerCase().includes(provLabel.toLowerCase())
73
+ return alreadyHasProv || !provLabel ? rawModelName : `${provLabel} · ${rawModelName}`
74
+ }
75
+
76
+ function detectUnkeyedProviders(runtime, llm) {
77
+ const missing = new Set()
78
+ if (!runtime || !llm || typeof llm.listConfigurableProviders !== 'function') return missing
79
+ try {
80
+ const settings = runtime.get('settings')
81
+ const credentials = runtime.get('credentials')
82
+ if (!settings || !credentials || typeof settings.describe !== 'function') return missing
83
+ const sections = new Map((settings.describe({ redactSecrets: true }) || []).map(s => [s.ns, s]))
84
+ for (const entry of (llm.listConfigurableProviders() || [])) {
85
+ if (!entry || !entry.provider || !Array.isArray(entry.settingsPath)) continue
86
+ const sec = sections.get(entry.settingsNs)
87
+ const profile = entry.settingsPath.reduce((v, k) => v?.[k], sec?.value)
88
+ const keyEnv = typeof profile?.apiKeyEnv === 'string' && profile.apiKeyEnv ? profile.apiKeyEnv : ''
89
+ if (keyEnv && process.env[keyEnv] === '') {
90
+ // Explicit empty env hint; async describe handled below when needed
91
+ }
92
+ }
93
+ } catch {}
94
+ return missing
95
+ }
96
+
97
+ async function detectMissingProviderKeysAsync(runtime, llm, providerIds) {
98
+ const missing = detectUnkeyedProviders(runtime, llm)
99
+ if (!runtime || !llm || typeof llm.listConfigurableProviders !== 'function') return missing
100
+ try {
101
+ const settings = runtime.get('settings')
102
+ const credentials = runtime.get('credentials')
103
+ if (!settings || !credentials || typeof settings.describe !== 'function' || typeof credentials.describe !== 'function') {
104
+ return missing
105
+ }
106
+ const sections = new Map((settings.describe({ redactSecrets: true }) || []).map(s => [s.ns, s]))
107
+ const keyRefs = new Map()
108
+ for (const entry of (llm.listConfigurableProviders() || [])) {
109
+ if (!entry || !entry.provider || !Array.isArray(entry.settingsPath)) continue
110
+ const sec = sections.get(entry.settingsNs)
111
+ const profile = entry.settingsPath.reduce((v, k) => v?.[k], sec?.value)
112
+ if (typeof profile?.apiKeyEnv === 'string' && profile.apiKeyEnv) {
113
+ keyRefs.set(entry.provider, profile.apiKeyEnv)
114
+ }
115
+ }
116
+ await Promise.all(
117
+ providerIds
118
+ .filter(id => keyRefs.has(id))
119
+ .map(async (id) => {
120
+ const info = await Promise.resolve(credentials.describe(keyRefs.get(id))).catch(() => undefined)
121
+ if (info && info.configured === false) missing.add(id)
122
+ }),
123
+ )
124
+ } catch {}
125
+ return missing
126
+ }
127
+
128
+ function extractAgentModelRef(agent) {
129
+ if (!agent) return ''
130
+ try {
131
+ const sess = agent.session
132
+ const header = sess && typeof sess.requestHeader === 'function'
133
+ ? sess.requestHeader()
134
+ : (sess && sess.header ? sess.header : null)
135
+ const candidates = [
136
+ header && header.config,
137
+ header,
138
+ sess && sess.config,
139
+ sess && sess.modelSelection,
140
+ agent.modelSelection,
141
+ agent.config,
142
+ ]
143
+ for (const cfg of candidates) {
144
+ if (cfg && typeof cfg.provider === 'string' && cfg.provider && typeof cfg.model === 'string' && cfg.model) {
145
+ if (!cfg.provider.startsWith('local-agent')) {
146
+ return `${cfg.provider}/${cfg.model}`
147
+ }
148
+ }
149
+ }
150
+ } catch {}
151
+ return ''
152
+ }
153
+
154
+ function sessionModelRef(deps, sessionId) {
155
+ if (!deps || !deps.runtime) return ''
156
+ try {
157
+ const agents = deps.runtime.get('agents')
158
+ if (!agents || typeof agents.get !== 'function') return ''
159
+ if (sessionId) {
160
+ const direct = extractAgentModelRef(agents.get(sessionId))
161
+ if (direct) return direct
162
+ const rootId = deps.resolveRootSessionId ? deps.resolveRootSessionId(sessionId) : ''
163
+ if (rootId && rootId !== sessionId) {
164
+ const rootRef = extractAgentModelRef(agents.get(rootId))
165
+ if (rootRef) return rootRef
166
+ }
167
+ }
168
+ } catch {}
169
+ return ''
170
+ }
171
+
172
+ function profileDefaultModelRef(runtime) {
173
+ if (!runtime) return ''
174
+ try {
175
+ const sel = runtime.get('agentDefaultModel')?.currentSelection?.()
176
+ if (sel && sel.provider && sel.model && !String(sel.provider).startsWith('local-agent')) {
177
+ return `${sel.provider}/${sel.model}`
178
+ }
179
+ } catch {}
180
+ return ''
181
+ }
182
+
183
+ export async function listChiefModels(deps, fresh = false, sessionId = '') {
184
+ const llm = deps.getLlm ? deps.getLlm() : null
185
+ const runtime = deps.runtime
186
+ const state = loadChiefState()
187
+ const rootSid = (deps && deps.resolveRootSessionId ? deps.resolveRootSessionId(sessionId) : '') || String(sessionId || '').trim()
188
+ const effCfg = resolveWorkspaceConfig(state, rootSid)
189
+ let selectedRef = effCfg.modelRef || ''
190
+
191
+ const leadRef = sessionModelRef(deps, sessionId)
192
+ const profileDefaultRef = profileDefaultModelRef(runtime)
193
+ const followRef = leadRef || profileDefaultRef || ''
194
+
195
+ if (!llm || typeof llm.listProviders !== 'function' || typeof llm.listModels !== 'function') {
196
+ const resolvedRef = selectedRef || followRef || ''
197
+ return {
198
+ available: false,
199
+ selectedRef,
200
+ followRef,
201
+ followLabel: formatChiefModelLabel(followRef, [], [], false),
202
+ defaultRef: followRef,
203
+ resolvedRef,
204
+ resolvedLabel: formatChiefModelLabel(resolvedRef, [], [], true),
205
+ models: [],
206
+ }
207
+ }
208
+
209
+ const rawProviders = (llm.listProviders() || []).filter(p => p && p.id && !String(p.id).startsWith('local-agent'))
210
+ const registeredProviderIds = new Set(rawProviders.map(p => String(p.id)))
211
+
212
+ // Auto-heal stale selectedRef if its provider was uninstalled/removed (e.g. legacy `opendesign/...`)
213
+ if (selectedRef) {
214
+ const parsedSel = splitModelRef(selectedRef)
215
+ if (!parsedSel || (registeredProviderIds.size > 0 && !registeredProviderIds.has(parsedSel.provider))) {
216
+ selectedRef = ''
217
+ mutateChiefState((s) => {
218
+ s.config.modelRef = ''
219
+ })
220
+ }
221
+ }
222
+
223
+ let items = []
224
+ if (!fresh && modelCatalogCache.items.length > 0 && Date.now() - modelCatalogCache.at < CATALOG_TTL_MS) {
225
+ items = modelCatalogCache.items.slice()
226
+ } else {
227
+ try {
228
+ const missingKeys = await detectMissingProviderKeysAsync(runtime, llm, [...registeredProviderIds])
229
+ for (const provider of rawProviders) {
230
+ const pid = String(provider.id)
231
+ const pname = String(provider.name || provider.label || pid)
232
+ const needsKey = missingKeys.has(pid)
233
+ let models = []
234
+ try {
235
+ models = (await llm.listModels(pid)) || []
236
+ } catch {
237
+ continue
238
+ }
239
+ for (const m of models) {
240
+ if (!m || !m.id) continue
241
+ items.push({
242
+ ref: `${pid}/${m.id}`,
243
+ provider: pid,
244
+ providerName: pname,
245
+ id: String(m.id),
246
+ name: String(m.name || m.label || m.id),
247
+ ...(needsKey ? { needsKey: true } : {}),
248
+ })
249
+ }
250
+ }
251
+ } catch {}
252
+ modelCatalogCache = { at: Date.now(), items: items.slice() }
253
+ }
254
+
255
+ // Ensure the active main-conversation model (`followRef`) & `profileDefaultRef` appear in the catalog
256
+ // even when a dynamic provider (e.g. Command Code) only returned a subset in `listModels()`!
257
+ const ensureListed = async (refToEnsure, markCurrent = false) => {
258
+ const parsed = splitModelRef(refToEnsure)
259
+ if (!parsed || !registeredProviderIds.has(parsed.provider)) return
260
+ const existing = items.find(it => it.ref === parsed.ref)
261
+ if (existing) {
262
+ if (markCurrent) existing.isCurrentLead = true
263
+ return
264
+ }
265
+ const provObj = rawProviders.find(p => String(p.id) === parsed.provider)
266
+ const pname = String((provObj && (provObj.name || provObj.label)) || parsed.provider)
267
+ let displayName = stripVendorPrefixFromModelId(parsed.model)
268
+ try {
269
+ if (typeof llm.resolveModelInfo === 'function') {
270
+ const info = await llm.resolveModelInfo(parsed.provider, parsed.model)
271
+ if (info && typeof info.name === 'string' && info.name.trim()) {
272
+ displayName = info.name.trim()
273
+ }
274
+ }
275
+ } catch {}
276
+ const entry = {
277
+ ref: parsed.ref,
278
+ provider: parsed.provider,
279
+ providerName: pname,
280
+ id: parsed.model,
281
+ name: displayName,
282
+ ...(markCurrent ? { isCurrentLead: true } : {}),
283
+ }
284
+ // Insert at the top of its provider group so it is immediately visible without scrolling
285
+ const firstProvIdx = items.findIndex(it => it.provider === parsed.provider)
286
+ if (firstProvIdx >= 0) items.splice(firstProvIdx, 0, entry)
287
+ else items.unshift(entry)
288
+ }
289
+
290
+ await ensureListed(followRef, true)
291
+ if (profileDefaultRef && profileDefaultRef !== followRef) {
292
+ await ensureListed(profileDefaultRef, false)
293
+ }
294
+ if (selectedRef && selectedRef !== followRef && selectedRef !== profileDefaultRef) {
295
+ await ensureListed(selectedRef, false)
296
+ }
297
+
298
+ const picked = pickRouteFromItems(items, selectedRef, followRef, deps, sessionId, registeredProviderIds)
299
+ const resolvedRef = (picked && picked.ref) || selectedRef || followRef || ''
300
+ return {
301
+ available: items.length > 0 || Boolean(resolvedRef),
302
+ selectedRef,
303
+ followRef,
304
+ followLabel: formatChiefModelLabel(followRef, items, rawProviders, false),
305
+ defaultRef: followRef,
306
+ resolvedRef,
307
+ resolvedLabel: formatChiefModelLabel(resolvedRef, items, rawProviders, true),
308
+ models: items,
309
+ }
310
+ }
311
+
312
+ const unusableRouteUntil = new Map()
313
+
314
+ function isRouteCoolingDown(ref, provider) {
315
+ const now = Date.now()
316
+ if (ref && (unusableRouteUntil.get(ref) || 0) > now) return true
317
+ if (provider && (unusableRouteUntil.get('prov:' + provider) || 0) > now) return true
318
+ return false
319
+ }
320
+
321
+ function markRouteFailure(route, err) {
322
+ if (!route) return
323
+ const msg = String(err && err.message ? err.message : err || '')
324
+ const code = String(err && (err.code || err.failure?.code) ? (err.code || err.failure?.code) : '')
325
+ // Missing API key / auth error / hosted loop engine cannot serve auxiliary completions
326
+ if (/api[_ -]?key|unauthorized|authentication|401|403|not configured|missing key|no key|credentials/i.test(msg) || /AUTH|KEY/i.test(code)) {
327
+ unusableRouteUntil.set('prov:' + route.provider, Date.now() + 90_000)
328
+ } else if (/HOSTED_ENGINE_ROUTE|not a model endpoint|no adapter/i.test(msg) || code === 'HOSTED_ENGINE_ROUTE') {
329
+ unusableRouteUntil.set(route.ref, Date.now() + 300_000)
330
+ } else {
331
+ unusableRouteUntil.set(route.ref, Date.now() + 20_000)
332
+ }
333
+ }
334
+
335
+ function buildRouteQueue(items, selectedRef, followRef, deps, sessionId, registeredProviderIds = null) {
336
+ const seen = new Set()
337
+ const out = []
338
+ const isProviderRegistered = (prov) => {
339
+ if (!prov || String(prov).startsWith('local-agent')) return false
340
+ if (registeredProviderIds && registeredProviderIds.size > 0) return registeredProviderIds.has(prov)
341
+ return true
342
+ }
343
+ const pushCand = (cand, allowUnlisted = false) => {
344
+ const parsed = splitModelRef(cand)
345
+ if (!parsed || !isProviderRegistered(parsed.provider)) return
346
+ if (seen.has(parsed.ref)) return
347
+ if (!allowUnlisted && items.length > 0 && !items.some(it => it.ref === parsed.ref && !it.needsKey)) return
348
+ seen.add(parsed.ref)
349
+ out.push(parsed)
350
+ }
351
+
352
+ const liveLeadRef = sessionModelRef(deps, sessionId)
353
+ const primaryFollow = liveLeadRef || followRef || profileDefaultModelRef(deps && deps.runtime)
354
+
355
+ if (selectedRef) {
356
+ // Explicit user-pinned Chief model: try it first, then same-provider models, then main session model
357
+ pushCand(selectedRef, true)
358
+ const selParsed = splitModelRef(selectedRef)
359
+ if (selParsed) {
360
+ for (const it of items) {
361
+ if (it && it.provider === selParsed.provider && !it.needsKey) pushCand(it.ref, false)
362
+ if (out.length >= 3) break
363
+ }
364
+ }
365
+ if (primaryFollow) pushCand(primaryFollow, true)
366
+ } else if (primaryFollow) {
367
+ // Auto-follow main conversation mode:
368
+ // MUST use the exact provider/model of the main conversation (e.g. `commandcode/deepseek/deepseek-v4.1-flash`)
369
+ // and NEVER silently jump across to a paid official API provider (like `deepseek/...`)!
370
+ pushCand(primaryFollow, true)
371
+ if (followRef && followRef !== primaryFollow) pushCand(followRef, true)
372
+ const followParsed = splitModelRef(primaryFollow)
373
+ if (followParsed) {
374
+ for (const it of items) {
375
+ if (it && it.provider === followParsed.provider && !it.needsKey) pushCand(it.ref, false)
376
+ if (out.length >= 4) break
377
+ }
378
+ }
379
+ } else {
380
+ // Only when neither a pinned model nor any main/default model exists at all: pick keyed catalog models
381
+ const seenProviders = new Set()
382
+ for (const it of items) {
383
+ if (!it || !it.ref || it.needsKey || seenProviders.has(it.provider)) continue
384
+ seenProviders.add(it.provider)
385
+ pushCand(it.ref, false)
386
+ }
387
+ }
388
+
389
+ // Prefer routes not currently cooling down, keeping cooled ones only as last-resort tail
390
+ const healthy = out.filter(r => !isRouteCoolingDown(r.ref, r.provider))
391
+ const cooled = out.filter(r => isRouteCoolingDown(r.ref, r.provider))
392
+ return [...healthy, ...cooled]
393
+ }
394
+
395
+ function pickRouteFromItems(items, selectedRef, followRef, deps, sessionId, registeredProviderIds = null) {
396
+ const queue = buildRouteQueue(items, selectedRef, followRef, deps, sessionId, registeredProviderIds)
397
+ return queue[0] || null
398
+ }
399
+
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 {}
415
+
416
+ const targetBudget = Math.max(Number(maxTokens) || 0, SUMMARY_MAX_TOKENS)
417
+ const firstBudget = modelDefaultMaxTokens ? Math.min(modelDefaultMaxTokens, targetBudget) : targetBudget
418
+ const retryBudget = firstBudget < targetBudget ? targetBudget : SUMMARY_RETRY_MAX_TOKENS
419
+ const budgets = retryBudget > firstBudget ? [firstBudget, retryBudget] : [firstBudget]
420
+ const safeImages = Array.isArray(imageBlocks) ? imageBlocks : []
421
+
422
+ const runOnce = async (budget) => {
423
+ const userContent = [...safeImages, { type: 'text', text: userText || '请查看附图并回答。' }]
424
+ const request = {
425
+ provider: route.provider,
426
+ model: route.model,
427
+ ...(sessionId ? { sessionId } : {}),
428
+ ...(effort !== undefined ? { reasoningEffort: effort } : {}),
429
+ system,
430
+ messages: [{ role: 'user', content: userContent }],
431
+ maxTokens: budget,
432
+ signal: AbortSignal.timeout(MODEL_TIMEOUT_MS),
433
+ }
434
+
435
+ const dshLlm = await loadDshLlm()
436
+ if (dshLlm && typeof dshLlm.BlockAssembler === 'function') {
437
+ const assembler = new dshLlm.BlockAssembler()
438
+ for await (const chunk of llm.stream(request)) {
439
+ assembler.push(chunk)
440
+ }
441
+ const finish = assembler.finish
442
+ if (finish && (finish.kind === 'error' || finish.kind === 'aborted')) {
443
+ const err = new Error(finish.failure?.message || finish.failure?.code || finish.kind || 'LLM request error')
444
+ err.failure = finish.failure
445
+ throw err
446
+ }
447
+ const text = assembler
448
+ .blocks()
449
+ .filter(b => b && b.type === 'text' && typeof b.text === 'string')
450
+ .map(b => b.text)
451
+ .join('\n')
452
+ .trim()
453
+ return { text, finishKind: finish?.kind || '', modelRef: route.ref }
454
+ }
455
+
456
+ // Pure standalone fallback stream reader (for unit tests / lightweight hosts)
457
+ const parts = []
458
+ let finishKind = ''
459
+ for await (const chunk of llm.stream(request)) {
460
+ if (!chunk || typeof chunk !== 'object') continue
461
+ if (chunk.type === 'text-delta' || chunk.type === 'text_delta') {
462
+ if (typeof chunk.text === 'string') parts.push(chunk.text)
463
+ else if (typeof chunk.delta === 'string') parts.push(chunk.delta)
464
+ } else if (chunk.type === 'text' && typeof chunk.text === 'string') parts.push(chunk.text)
465
+ else if (typeof chunk.content === 'string') parts.push(chunk.content)
466
+ else if (chunk.type === 'finish') {
467
+ finishKind = chunk.reason?.kind || chunk.finishReason || ''
468
+ if (finishKind === 'error') {
469
+ const err = new Error(chunk.reason?.failure?.message || 'LLM stream error')
470
+ err.failure = chunk.reason?.failure
471
+ throw err
472
+ }
473
+ }
474
+ }
475
+ return { text: parts.join('').trim(), finishKind, modelRef: route.ref }
476
+ }
477
+
478
+ let lastRes = { text: '', finishKind: '', modelRef: route.ref }
479
+ for (let i = 0; i < budgets.length; i += 1) {
480
+ const res = await runOnce(budgets[i])
481
+ lastRes = res
482
+ if (res.text && res.finishKind !== 'max-tokens') {
483
+ return { text: res.text, modelRef: res.modelRef }
484
+ }
485
+ }
486
+ return { text: lastRes.text, modelRef: lastRes.modelRef }
487
+ }
488
+
489
+ async function filterVisionRoutes(llm, queue) {
490
+ const visionRoutes = []
491
+ const unknownRoutes = []
492
+ for (const route of queue) {
493
+ let info = null
494
+ try {
495
+ if (typeof llm.resolveModelInfo === 'function') {
496
+ info = await Promise.resolve(llm.resolveModelInfo(route.provider, route.model))
497
+ }
498
+ } catch {}
499
+ const mods = Array.isArray(info?.inputModalities)
500
+ ? info.inputModalities
501
+ : (Array.isArray(route.inputModalities) ? route.inputModalities : null)
502
+ if (mods) {
503
+ if (mods.includes('image')) visionRoutes.push(route)
504
+ } else {
505
+ unknownRoutes.push(route)
506
+ }
507
+ }
508
+ return [...visionRoutes, ...unknownRoutes]
509
+ }
510
+
511
+ export async function callChiefLlm(deps, {
512
+ sessionId = '',
513
+ system,
514
+ prompt,
515
+ maxTokens = SUMMARY_MAX_TOKENS,
516
+ imageBlocks = [],
517
+ allowTextFallbackOnVisionMiss = false,
518
+ }) {
519
+ const safeImages = Array.isArray(imageBlocks) ? imageBlocks : []
520
+ if (deps && typeof deps.callLlmOverride === 'function') {
521
+ return deps.callLlmOverride({ sessionId, system, prompt, maxTokens, imageBlocks: safeImages, allowTextFallbackOnVisionMiss })
522
+ }
523
+ const llm = deps.getLlm ? deps.getLlm() : null
524
+ if (!llm || typeof llm.stream !== 'function') {
525
+ return { ok: false, error: '当前环境未挂载 LLM 服务' }
526
+ }
527
+ const catalog = await listChiefModels(deps, false, sessionId)
528
+ const regSet = typeof llm.listProviders === 'function'
529
+ ? new Set((llm.listProviders() || []).filter(p => p && p.id && !String(p.id).startsWith('local-agent')).map(p => String(p.id)))
530
+ : null
531
+ const baseQueue = buildRouteQueue(catalog.models, catalog.selectedRef, catalog.followRef || catalog.defaultRef, deps, sessionId, regSet)
532
+ 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' }
534
+ }
535
+
536
+ let queue = baseQueue
537
+ let sendImages = safeImages
538
+ let degradedToTextOnly = false
539
+ if (safeImages.length > 0) {
540
+ const visionQueue = await filterVisionRoutes(llm, baseQueue)
541
+ if (visionQueue.length > 0) {
542
+ queue = visionQueue
543
+ } else if (allowTextFallbackOnVisionMiss) {
544
+ queue = baseQueue
545
+ sendImages = []
546
+ degradedToTextOnly = true
547
+ } else {
548
+ return {
549
+ ok: false,
550
+ 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。',
551
+ }
552
+ }
553
+ }
554
+
555
+ let lastError = '模型返回内容为空'
556
+ for (const route of queue.slice(0, 4)) {
557
+ 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 }
560
+ lastError = `${route.ref} 返回内容为空`
561
+ markRouteFailure(route, new Error(lastError))
562
+ } catch (err) {
563
+ lastError = String(err && err.message ? err.message : err)
564
+ if (sendImages.length > 0 && allowTextFallbackOnVisionMiss) {
565
+ 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 }
568
+ } catch {}
569
+ }
570
+ markRouteFailure(route, err)
571
+ }
572
+ }
573
+ return { ok: false, error: lastError, degradedToTextOnly }
574
+ }
575
+
576
+ /* -------------------------------------------------------------------------- */
577
+ /* Peer Discovery & Cross-Conversation Delivery ("递话 / 插话") */
578
+ /* -------------------------------------------------------------------------- */