dsh-vertex-argon 2.2.6 → 2.2.8

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/CHANGELOG.md CHANGED
@@ -2,6 +2,32 @@
2
2
 
3
3
  All notable changes to **Gemini 4 Adapter** (`dsh-vertex-argon`) are documented in this file.
4
4
 
5
+ ## [2.2.8] - 2026-11-19
6
+
7
+ ### Fixed
8
+ - **Strict Candidate Part Prototype, Empty-Key & Same-Part Dual Signature Rejection (`google-signatures.js`)**:
9
+ - Rejected dual `thoughtSignature` + `thought_signature` and dual `id` + `call_id` unconditionally even when values match, checked invalid call keys by array length (catching `""` keys), and enforced plain-object (`Object.prototype` / `null`) checks on parts, `functionCall` payloads, and `call.args`.
10
+ - **Indexed Snapshot Rollback for Unindexed Interactions Deltas (`interactions.js`)**:
11
+ - Captured an indexed state snapshot after each event carrying a valid `event_id` and restored it before reconnecting with `last_event_id`, preventing un-indexed `function_call` argument deltas prior to a stream drop from duplicating on replay while preserving cursorless retry bounds.
12
+ - **Request-Wide Committed Prefix & Accountable Budget Guards on Thinking-Dropout Heal (`generate-content.js`)**:
13
+ - Tracked request-wide `anyCommittedVisibleOrTool` across hops so synthetic (`!hopToken`) `thinking-dropout-heal` never replays without prior committed text/tool calls, and required accountable non-null output token usage before consuming remaining `maxOutputTokens` on synthetic heal.
14
+ - **Array Sparse-Hole / Symbol Deep-Freeze & Explicit `functionCallingConfig.mode` Preservation (`adapter.js`)**:
15
+ - Updated `deepFreezeClone` to preserve sparse array holes, custom array properties, and array Symbol keys via `new Array(length)` + `Reflect.ownKeys`, and preserved caller-specified `functionCallingConfig.mode` (such as `'NONE'` and `'ANY'`) instead of overwriting with `'VALIDATED'`.
16
+
17
+ ## [2.2.7] - 2026-11-19
18
+
19
+ ### Fixed
20
+ - **Same-Chunk Part Splitting & Dual-Key Field Validation (`google-signatures.js`)**:
21
+ - Split subsequent `text` / `thinking` parts within the same candidate `parts` array even when neither carries a `thoughtSignature`, aligning streaming and non-streaming block counts 1:1.
22
+ - Enforced strict individual validation on `thoughtSignature` / `thought_signature` and `functionCall` / `function_call` (rejecting empty/`null`/conflicting dual spellings, missing or empty `functionCall.name`, unsupported keys outside `name`/`args`/`id`/`call_id`, non-plain-object `args`, and conflicting `id`/`call_id`).
23
+ - **Unindexed Visible Interactions Frame Guard (`interactions.js`)**:
24
+ - Tracked un-indexed visible output (`unindexedVisibleOutput`) even after an earlier `event_id`, failing closed instead of replaying from an older cursor unless a subsequent `event_id` advances `lastEventId`.
25
+ - **Pre-Handler Interactions Reasoning Validation & Recursive Descriptor Freeze (`adapter.js`)**:
26
+ - Validated `resolveGemini4ReasoningLevel` inside `ArgonAdapter.prepareCall` before invoking custom `interactionHandler` instances, returning an in-band `{ type: 'finish', reason: { kind: 'error' } }` frame on invalid or conflicting options.
27
+ - Upgraded `freezeDescriptor` to recursive `deepFreezeClone` across nested objects, arrays, and symbol properties.
28
+ - **Thinking-Dropout Heal Token Budget Deduction (`generate-content.js`)**:
29
+ - Evaluated Hop 1 usage before deciding `hideStop` / `canHeal()` and deducted Hop 1 spent candidate + thought tokens from `maxOutputTokens` / `max_output_tokens` across both `request.body` and cloned `httpOptions.extraBody` on synthetic (`!hopToken`) replay hops (refusing synthetic healing when no budget remains), while leaving checkpointed `hopToken` requests untouched.
30
+
5
31
  ## [2.2.6] - 2026-11-19
6
32
 
7
33
  ### Fixed
package/README.md CHANGED
@@ -4,7 +4,7 @@
4
4
 
5
5
  **面向 [DeepSeek Harness](<https://github.com/deepseek-ai/deepseek-harness>) 的 Google 官方双通道(Vertex AI + Gemini Developer API)Gemini 4 适配器,原生支持 Argon、Carbon、Hydrogen 与 Neon 八模型路由、GenerateContent / Interactions 双协议、65K+ 多跳长解码与请求级诊断。**
6
6
 
7
- - **当前版本**:`2.2.6`([CHANGELOG.md](<CHANGELOG.md>))
7
+ - **当前版本**:`2.2.8`([CHANGELOG.md](<CHANGELOG.md>))
8
8
  - **npm 包名**:`dsh-vertex-argon`(保留既有包名、官方通道路由与本地存储域 ID,官方通道用户可直接原地升级;`2.0.0` 已移除 `1.6.x` 非官方反代通道)
9
9
  - **运行要求**:DSH `>=0.2.0-rc.2`,Node.js `^22.19.0 || >=24.0.0`(见 [package.json](<package.json>))
10
10
 
package/adapter.js CHANGED
@@ -605,6 +605,13 @@ export function gemini4StreamSimple(
605
605
  : (cleanConfig.max_output_tokens !== undefined ? cleanConfig.max_output_tokens : (effectiveMaxTokens ?? outboundCeiling))
606
606
  delete cleanConfig.max_output_tokens
607
607
  const toolMode = shouldFallbackToAutoToolMode(retryKey) ? 'AUTO' : 'VALIDATED'
608
+ const rawExistingMode = cleanConfig.toolConfig?.functionCallingConfig?.mode
609
+ const existingModeUpper = typeof rawExistingMode === 'string' ? rawExistingMode.trim().toUpperCase() : ''
610
+ const preserveCallerMode =
611
+ existingModeUpper === 'NONE' ||
612
+ existingModeUpper === 'ANY' ||
613
+ (existingModeUpper !== '' && existingModeUpper !== 'VALIDATED')
614
+ const effectiveToolMode = preserveCallerMode ? rawExistingMode : toolMode
608
615
  return {
609
616
  ...body,
610
617
  model: upstreamModelId,
@@ -614,7 +621,7 @@ export function gemini4StreamSimple(
614
621
  ...(cleanConfig.tools?.some(tool => tool.functionDeclarations?.length) ? {
615
622
  toolConfig: {
616
623
  ...cleanConfig.toolConfig,
617
- functionCallingConfig: { ...cleanConfig.toolConfig?.functionCallingConfig, mode: toolMode },
624
+ functionCallingConfig: { ...cleanConfig.toolConfig?.functionCallingConfig, mode: effectiveToolMode },
618
625
  },
619
626
  } : {}),
620
627
  automaticContinuation: false,
@@ -852,24 +859,33 @@ export function resolveConfigValues(raw = {}) {
852
859
  }
853
860
  }
854
861
 
855
- function freezeDescriptor(descriptor) {
856
- if (!descriptor || typeof descriptor !== 'object') return descriptor
857
- const copy = {
858
- ...descriptor,
859
- ...(Array.isArray(descriptor.input) ? { input: Object.freeze([...descriptor.input]) } : {}),
860
- ...(descriptor.thinkingLevelMap && typeof descriptor.thinkingLevelMap === 'object'
861
- ? { thinkingLevelMap: Object.freeze({ ...descriptor.thinkingLevelMap }) }
862
- : {}),
863
- ...(descriptor.cost && typeof descriptor.cost === 'object'
864
- ? { cost: Object.freeze({ ...descriptor.cost }) }
865
- : {}),
862
+ export function deepFreezeClone(value, seen = new WeakMap()) {
863
+ if (!value || typeof value !== 'object') return value
864
+ if (seen.has(value)) return seen.get(value)
865
+ if (Array.isArray(value)) {
866
+ const arr = new Array(value.length)
867
+ seen.set(value, arr)
868
+ for (const key of Reflect.ownKeys(value)) {
869
+ if (key === 'length') continue
870
+ arr[key] = deepFreezeClone(value[key], seen)
871
+ }
872
+ return Object.freeze(arr)
866
873
  }
867
- for (const sym of Object.getOwnPropertySymbols(descriptor)) {
868
- copy[sym] = descriptor[sym]
874
+ const copy = {}
875
+ seen.set(value, copy)
876
+ for (const key of Object.keys(value)) {
877
+ copy[key] = deepFreezeClone(value[key], seen)
878
+ }
879
+ for (const sym of Object.getOwnPropertySymbols(value)) {
880
+ copy[sym] = deepFreezeClone(value[sym], seen)
869
881
  }
870
882
  return Object.freeze(copy)
871
883
  }
872
884
 
885
+ function freezeDescriptor(descriptor) {
886
+ return deepFreezeClone(descriptor)
887
+ }
888
+
873
889
  /** Serve the official Vertex AI and Gemini API models in one plugin group. */
874
890
  export function argonProvider(nativeVertex, models, defaultLevel = undefined, nativeGoogle = googleProvider(), longDecode = true, recordDiagnostic, serviceConfig = {}, responseHub, structuredOutput) {
875
891
  const frozenModels = (Array.isArray(models) ? models : [models]).map(freezeDescriptor)
@@ -978,6 +994,17 @@ export class ArgonAdapter extends LlmAdapter {
978
994
  stream: options => {
979
995
  const pinned = pinPreparedOptions(options, provider, model)
980
996
  assertGemini4MaxTokens(pinned, resolvedModel, cfg?.longDecode)
997
+ const defaultReasoning = cfg?.hasExplicitReasoning === false ? defaultReasoningForModel(model) : (cfg?.reasoning ?? defaultReasoningForModel(model))
998
+ const check = resolveGemini4ReasoningLevel(pinned, model, defaultReasoning)
999
+ if (!check.ok) {
1000
+ const message = check.errorType === 'conflict'
1001
+ ? 'Conflicting reasoning and reasoningEffort options in Interactions request.'
1002
+ : (check.supportsXhigh ? 'Gemini 4 (Hydrogen / Neon) requires low, medium, high, or xhigh thinking.' : 'Argon requires low, medium, or high thinking.')
1003
+ const code = check.errorType === 'conflict' ? 'INVALID_REQUEST' : 'UNSUPPORTED_REASONING_EFFORT'
1004
+ return (async function* () {
1005
+ yield { type: 'finish', reason: { kind: 'error', failure: { code, message } } }
1006
+ })()
1007
+ }
981
1008
  return this.interactionHandler(syncGemini4ReasoningOptions(pinned, model), cfg)
982
1009
  },
983
1010
  }
@@ -222,6 +222,14 @@ export async function continueArgonRequestStream(apiClient, original, request, m
222
222
  return (async function* () {
223
223
  const baseBody = JSON.parse(request.body)
224
224
  let initialBody = request.body
225
+ const readMaxBudget = obj => {
226
+ for (const cfgObj of [obj?.generationConfig, obj?.generation_config]) {
227
+ const val = cfgObj?.maxOutputTokens ?? cfgObj?.max_output_tokens
228
+ if (Number.isSafeInteger(val) && val > 0) return val
229
+ }
230
+ return undefined
231
+ }
232
+ const initialMaxOutputTokens = readMaxBudget(httpOptions.extraBody) ?? readMaxBudget(baseBody)
225
233
  // Google documents a top-level service_tier. Apply after SDK conversion so
226
234
  // older converters cannot silently discard it; every hop uses this snapshot.
227
235
  if (!apiClient.isVertexAI?.() && meta.serviceTier !== undefined) {
@@ -243,6 +251,7 @@ export async function continueArgonRequestStream(apiClient, original, request, m
243
251
  let initialPrompt, initialCached
244
252
  let initialUsageLocked = false
245
253
  let thinkingDropoutHeals = 0
254
+ let anyCommittedVisibleOrTool = false
246
255
  const totals = Object.fromEntries(fields.map(key => [key, 0]))
247
256
  const billed = Object.fromEntries(fields.map(key => [key, 0]))
248
257
 
@@ -259,7 +268,22 @@ export async function continueArgonRequestStream(apiClient, original, request, m
259
268
  let blocked = false, terminal = false, streamSettled = false, usageAccounted = false
260
269
  let hopChunks = 0, hopSawThought = false, hopSawVisibleOrTool = false
261
270
  let thoughtParts = [], thoughtBytes = 0, suppressedStop = false, hopToken
262
- const canHeal = () => enabled && thinkingDropoutHeals < MAX_THINKING_DROPOUT_HEALS && hopSawThought && !hopSawVisibleOrTool && !blocked && finishReason === 'STOP'
271
+ const accountableHopOutputSpent = () => {
272
+ if (!latestUsage || latestUsage.candidatesTokenCount === null || latestUsage.thoughtsTokenCount === null || latestUsage.totalTokenCount === null) return null
273
+ const explicitThoughts = Number.isSafeInteger(latestUsage.thoughtsTokenCount)
274
+ const derivedFromTotal = Number.isSafeInteger(latestUsage.totalTokenCount) && Number.isSafeInteger(latestUsage.promptTokenCount)
275
+ ? Math.max(0, latestUsage.totalTokenCount - latestUsage.promptTokenCount - value(latestUsage, 'toolUsePromptTokenCount'))
276
+ : null
277
+ if (!explicitThoughts && derivedFromTotal === null) return null
278
+ return Math.max(value(latestUsage, 'candidatesTokenCount') + value(latestUsage, 'thoughtsTokenCount'), derivedFromTotal ?? 0)
279
+ }
280
+ const remainingHealBudget = () => {
281
+ if (initialMaxOutputTokens === undefined) return Infinity
282
+ const currentHopSpent = usageAccounted ? 0 : accountableHopOutputSpent()
283
+ if (currentHopSpent === null) return 0
284
+ return initialMaxOutputTokens - (billed.candidatesTokenCount + billed.thoughtsTokenCount + currentHopSpent)
285
+ }
286
+ const canHeal = () => enabled && thinkingDropoutHeals < MAX_THINKING_DROPOUT_HEALS && hopSawThought && !hopSawVisibleOrTool && !blocked && finishReason === 'STOP' && (Boolean(hopToken) || (!anyCommittedVisibleOrTool && !startToken && remainingHealBudget() > 0))
263
287
 
264
288
  const wrap = (raw, headers, usageSnapshot, hideStop = false) => {
265
289
  let out = raw
@@ -349,8 +373,9 @@ export async function continueArgonRequestStream(apiClient, original, request, m
349
373
  }
350
374
  const candidate = raw.candidates?.[0]
351
375
  const parts = candidate?.content?.parts ?? []
376
+ const chunkHasVisibleOrTool = parts.some(part => Boolean(part.functionCall || part.function_call || (!part.thought && typeof part.text === 'string' && part.text.length > 0)))
352
377
  hopSawThought ||= parts.some(part => part.thought === true)
353
- hopSawVisibleOrTool ||= parts.some(part => part.functionCall || (!part.thought && typeof part.text === 'string' && part.text.length > 0))
378
+ hopSawVisibleOrTool ||= chunkHasVisibleOrTool
354
379
  if (hopSawVisibleOrTool) { thoughtParts = []; thoughtBytes = 0 }
355
380
  else if (enabled && thinkingDropoutHeals < MAX_THINKING_DROPOUT_HEALS) {
356
381
  for (const part of parts) if (part.thought === true || part.thoughtSignature || part.thought_signature) {
@@ -368,24 +393,25 @@ export async function continueArgonRequestStream(apiClient, original, request, m
368
393
  // continuation token, a repeated finishReason or an empty thought frame
369
394
  // must never extend a stalled hop.
370
395
  watchdog.update(parts, Boolean(newCheckpoint || (candidate?.finishReason !== undefined && candidate.finishReason !== finishReason) || (!blocked && raw.promptFeedback?.blockReason)))
396
+ if (raw.usageMetadata && typeof raw.usageMetadata === 'object') latestUsage = readUsage(raw.usageMetadata, latestUsage)
371
397
  if (candidate?.finishReason !== undefined) finishReason = candidate.finishReason
372
398
  blocked ||= Boolean(raw.promptFeedback?.blockReason)
373
399
  terminal ||= blocked || (finishReason !== undefined && finishReason !== 'CONTINUATION')
374
400
  const hideStop = candidate?.finishReason === 'STOP' && canHeal()
375
401
  if (candidate?.finishReason !== undefined && candidate.finishReason !== 'CONTINUATION') suppressedStop = hideStop
376
402
  hopTracker?.onChunk({ status: 200, hasCheckpoint: hasToken, finishReason: candidate?.finishReason })
377
- if (raw.usageMetadata && typeof raw.usageMetadata === 'object') latestUsage = readUsage(raw.usageMetadata, latestUsage)
378
403
  if (!enabled || (!token && !newCheckpoint && !terminal && candidate?.finishReason !== 'CONTINUATION')) {
379
404
  refreshOriginalUsage()
380
405
  const outgoing = wrap(raw, chunk.headers, latestUsage, hideStop)
381
406
  committedUsage = latestUsage ? { ...latestUsage } : undefined
407
+ anyCommittedVisibleOrTool ||= chunkHasVisibleOrTool
382
408
  hopChunks++
383
409
  yield outgoing
384
410
  continue
385
411
  }
386
412
  pendingBytes += Buffer.byteLength(JSON.stringify(raw))
387
413
  if (pendingBytes > MAX_PENDING_BYTES) throw streamError('Argon uncommitted stream exceeded the bounded checkpoint buffer.', 'ARGON_BUFFER_LIMIT')
388
- pending.push({ raw, headers: chunk.headers, usage: latestUsage ? { ...latestUsage } : undefined, hideStop })
414
+ pending.push({ raw, headers: chunk.headers, usage: latestUsage ? { ...latestUsage } : undefined, hideStop, hasVisibleOrTool: chunkHasVisibleOrTool })
389
415
  pendingHasContent ||= parts.some(part => typeof part.text === 'string' ? part.text.length > 0 || Boolean(part.thoughtSignature || part.thought_signature) : Object.keys(part).length > 0)
390
416
  const finalMetadata = finishReason === 'CONTINUATION' && token && token !== startToken && !pendingHasContent
391
417
  if (newCheckpoint || terminal || finalMetadata) {
@@ -397,6 +423,7 @@ export async function continueArgonRequestStream(apiClient, original, request, m
397
423
  throwIfAborted(request.abortSignal)
398
424
  const outgoing = wrap(saved.raw, saved.headers, saved.usage, saved.hideStop)
399
425
  committedUsage = saved.usage
426
+ anyCommittedVisibleOrTool ||= saved.hasVisibleOrTool
400
427
  hopChunks++
401
428
  yield outgoing
402
429
  }
@@ -427,15 +454,37 @@ export async function continueArgonRequestStream(apiClient, original, request, m
427
454
  }
428
455
  if (terminal) {
429
456
  if (!failure && canHeal()) {
457
+ const priorBilledOutput = billed.candidatesTokenCount + billed.thoughtsTokenCount
458
+ const hopOutputSpent = accountableHopOutputSpent() ?? 0
430
459
  accountUsage()
431
460
  hopTracker?.finish({ finishReason, continued: true })
432
461
  if (hop === MAX_CONTINUATION_HOPS) throw streamError('Argon continuation hop limit reached without a terminal response.', 'ARGON_HOP_LIMIT')
433
462
  thinkingDropoutHeals++
434
- if (hopToken) token = hopToken
435
- else {
463
+ if (hopToken) {
464
+ token = hopToken
465
+ } else {
436
466
  token = undefined
467
+ const remainingBudget = initialMaxOutputTokens === undefined ? undefined : initialMaxOutputTokens - (priorBilledOutput + hopOutputSpent)
468
+ const withRemainingBudget = bodyObj => {
469
+ if (remainingBudget === undefined) return bodyObj
470
+ const updateCfg = cfgObj => {
471
+ if (!cfgObj || typeof cfgObj !== 'object' || Array.isArray(cfgObj)) return cfgObj
472
+ const nextCfg = { ...cfgObj }
473
+ if ('maxOutputTokens' in nextCfg) nextCfg.maxOutputTokens = remainingBudget
474
+ if ('max_output_tokens' in nextCfg) nextCfg.max_output_tokens = remainingBudget
475
+ return nextCfg
476
+ }
477
+ return {
478
+ ...bodyObj,
479
+ ...(bodyObj.generationConfig !== undefined ? { generationConfig: updateCfg(bodyObj.generationConfig) } : {}),
480
+ ...(bodyObj.generation_config !== undefined ? { generation_config: updateCfg(bodyObj.generation_config) } : {}),
481
+ }
482
+ }
483
+ if (httpOptions.extraBody && typeof httpOptions.extraBody === 'object' && !Array.isArray(httpOptions.extraBody)) {
484
+ httpOptions.extraBody = withRemainingBudget(httpOptions.extraBody)
485
+ }
437
486
  const { continuationToken: ignored, ...rest } = continuationBody
438
- continuationBody = { ...rest, contents: [...(rest.contents ?? []), { role: 'model', parts: thoughtParts }, { role: 'user', parts: [{ text: 'Continue and provide the final response or tool call.' }] }] }
487
+ continuationBody = withRemainingBudget({ ...rest, contents: [...(rest.contents ?? []), { role: 'model', parts: thoughtParts }, { role: 'user', parts: [{ text: 'Continue and provide the final response or tool call.' }] }] })
439
488
  }
440
489
  retriesWithoutProgress = 0; retryAttempt = 0; delayMs = 0
441
490
  trigger = 'thinking-dropout-heal'
@@ -1,6 +1,13 @@
1
1
  import { createHash, randomUUID } from 'node:crypto'
2
2
 
3
3
  const ALLOWED_PART_KEYS = new Set(['text', 'thought', 'thoughtSignature', 'thought_signature', 'functionCall', 'function_call'])
4
+ const ALLOWED_CALL_KEYS = new Set(['name', 'args', 'id', 'call_id'])
5
+
6
+ function isPlainRecord(value) {
7
+ if (!value || typeof value !== 'object' || Array.isArray(value)) return false
8
+ const proto = Object.getPrototypeOf(value)
9
+ return proto === Object.prototype || proto === null
10
+ }
4
11
 
5
12
  const PREFIX = '{"dshGoogleSignatureOnly":'
6
13
  const base64 = value => typeof value === 'string' && value.length > 0 && value.length % 4 === 0 && /^[A-Za-z0-9+/]+={0,2}$/.test(value)
@@ -164,7 +171,7 @@ export function createGoogleSignatureBridge(context, model) {
164
171
  seen = true
165
172
  currentFromThisChunk = false
166
173
  for (const part of chunk.candidates?.[0]?.content?.parts ?? []) {
167
- if (!part || typeof part !== 'object' || Array.isArray(part)) {
174
+ if (!isPlainRecord(part)) {
168
175
  close()
169
176
  throw failure('Google returned a malformed candidate part.')
170
177
  }
@@ -175,6 +182,18 @@ export function createGoogleSignatureBridge(context, model) {
175
182
  close()
176
183
  throw failure('Google returned an unsupported candidate part (' + unsupportedKeys.join(', ') + ') that cannot be replayed.')
177
184
  }
185
+ if (part.functionCall !== undefined && !isPlainRecord(part.functionCall)) {
186
+ close()
187
+ throw failure('Google returned a candidate part with an invalid functionCall payload.')
188
+ }
189
+ if (part.function_call !== undefined && !isPlainRecord(part.function_call)) {
190
+ close()
191
+ throw failure('Google returned a candidate part with an invalid functionCall payload.')
192
+ }
193
+ if (part.functionCall !== undefined && part.function_call !== undefined) {
194
+ close()
195
+ throw failure('Google returned a candidate part with both functionCall and function_call fields.')
196
+ }
178
197
  const call = part.functionCall ?? part.function_call
179
198
  const hasText = part.text !== undefined
180
199
  const hasCall = call !== undefined
@@ -186,9 +205,32 @@ export function createGoogleSignatureBridge(context, model) {
186
205
  close()
187
206
  throw failure('Google returned a candidate part with an invalid text field.')
188
207
  }
189
- if (hasCall && (!call || typeof call !== 'object' || Array.isArray(call))) {
190
- close()
191
- throw failure('Google returned a candidate part with an invalid functionCall payload.')
208
+ if (hasCall) {
209
+ const invalidCallKeys = Object.keys(call).filter(key => call[key] !== undefined && !ALLOWED_CALL_KEYS.has(key))
210
+ if (invalidCallKeys.length > 0) {
211
+ close()
212
+ throw failure('Google returned a candidate part with an unsupported functionCall field (' + JSON.stringify(invalidCallKeys[0]) + ').')
213
+ }
214
+ if (typeof call.name !== 'string' || !call.name) {
215
+ close()
216
+ throw failure('Google returned a candidate part with an invalid functionCall name.')
217
+ }
218
+ if (call.args !== undefined && !isPlainRecord(call.args)) {
219
+ close()
220
+ throw failure('Google returned a candidate part with an invalid functionCall args payload.')
221
+ }
222
+ if (call.id !== undefined && (typeof call.id !== 'string' || !call.id)) {
223
+ close()
224
+ throw failure('Google returned a candidate part with an invalid functionCall id.')
225
+ }
226
+ if (call.call_id !== undefined && (typeof call.call_id !== 'string' || !call.call_id)) {
227
+ close()
228
+ throw failure('Google returned a candidate part with an invalid functionCall call_id.')
229
+ }
230
+ if (call.id !== undefined && call.call_id !== undefined) {
231
+ close()
232
+ throw failure('Google returned a candidate part with both functionCall id and call_id fields.')
233
+ }
192
234
  }
193
235
  if (hasText && hasCall) {
194
236
  close()
@@ -200,12 +242,20 @@ export function createGoogleSignatureBridge(context, model) {
200
242
  }
201
243
  bytes += Buffer.byteLength(JSON.stringify(part))
202
244
  if (bytes > 64 * 1024 * 1024) throw failure('Google response exceeded the bounded signature replay buffer.')
203
- // A Part may serialize the signature in either wire spelling.
204
- const sig = part.thoughtSignature ?? part.thought_signature
205
- if (sig !== undefined && typeof sig !== 'string') {
245
+ // A Part may serialize the signature in either wire spelling, but never both on one Part.
246
+ if (part.thoughtSignature !== undefined && (typeof part.thoughtSignature !== 'string' || !part.thoughtSignature)) {
247
+ close()
248
+ throw failure('Google returned a candidate part with an invalid thoughtSignature.')
249
+ }
250
+ if (part.thought_signature !== undefined && (typeof part.thought_signature !== 'string' || !part.thought_signature)) {
206
251
  close()
207
252
  throw failure('Google returned a candidate part with an invalid thoughtSignature.')
208
253
  }
254
+ if (part.thoughtSignature !== undefined && part.thought_signature !== undefined) {
255
+ close()
256
+ throw failure('Google returned a candidate part with both thoughtSignature and thought_signature fields.')
257
+ }
258
+ const sig = part.thoughtSignature ?? part.thought_signature
209
259
  const onlySignature = typeof sig === 'string' && !hasText && !hasCall
210
260
  if (onlySignature) {
211
261
  if (!sig) {
@@ -225,13 +275,12 @@ export function createGoogleSignatureBridge(context, model) {
225
275
  const sigKey = type === 'text' ? 'textSignature' : 'thinkingSignature'
226
276
  const curText = type === 'text' ? current?.text : current?.thinking
227
277
  const curSig = current?.[sigKey]
228
- const sameChunkSigBoundary = currentFromThisChunk && (Boolean(curSig) || Boolean(sig))
229
278
  const crossChunkSigConflict =
230
279
  (Boolean(curSig) && Boolean(sig)) ||
231
280
  (Boolean(curSig) && curText === '') ||
232
281
  (Boolean(sig) && part.text === '')
233
- const splitSigned = sameChunkSigBoundary || crossChunkSigConflict
234
- if (!current || current.type !== type || splitSigned) {
282
+ const splitPart = currentFromThisChunk || crossChunkSigConflict
283
+ if (!current || current.type !== type || splitPart) {
235
284
  close(); current = type === 'text' ? { type, text: '' } : { type, thinking: '' }
236
285
  const index = start(current)
237
286
  emit({ type: type === 'text' ? 'text_start' : 'thinking_start', contentIndex: index })
package/interactions.js CHANGED
@@ -29,20 +29,24 @@ export async function* streamInteractions(options, settings) {
29
29
  const signal=options.signal, transport=settings.transport
30
30
  const steps=new Map(), blocks=new Map(), seenEvents=new Set(), emittedToolIds=new Set()
31
31
  let exitReason='consumer-abandoned'
32
- let blockIndex=0, bytes=0, interactionId, lastEventId, status, usage, cancelled=false, finishedNormally=false, emitted=false, completedEvent=false, progressVersion=0
32
+ let blockIndex=0, bytes=0, interactionId, lastEventId, status, usage, cancelled=false, finishedNormally=false, emitted=false, unindexedVisibleOutput=false, completedEvent=false, progressVersion=0, indexedSnapshot
33
+ const cloneStepState=state=>({...state,...(state.argsValue!==undefined?{argsValue:structuredClone(state.argsValue)}:{}),...(state.args!==undefined?{args:structuredClone(state.args)}:{})})
34
+ const captureSnapshot=()=>({steps:new Map([...steps.entries()].map(([k,v])=>[k,cloneStepState(v)])),blocks:new Map([...blocks.entries()].map(([k,v])=>[k,structuredClone(v)])),blockIndex,bytes,status,usage:usage?structuredClone(usage):undefined,completedEvent})
35
+ const restoreSnapshot=snap=>{steps.clear();for(const[k,v]of snap.steps.entries())steps.set(k,cloneStepState(v));blocks.clear();for(const[k,v]of snap.blocks.entries())blocks.set(k,structuredClone(v));blockIndex=snap.blockIndex;bytes=snap.bytes;status=snap.status;usage=snap.usage?structuredClone(snap.usage):undefined;completedEvent=snap.completedEvent}
33
36
  const observeUsage=event=>{const u=event.interaction?.usage??event.usage??event.metadata?.total_usage;if(validObject(u)){const before=(usage?.total_output_tokens??0)+(usage?.total_thought_tokens??0);usage={...usage,...u};if((usage.total_output_tokens??0)+(usage.total_thought_tokens??0)>before)progressVersion++}}
34
37
  const checkBytes=text=>{bytes+=Buffer.byteLength(text??'');if(bytes>MAX_BYTES)throw protocolError('Interactions output exceeded the bounded response buffer.')}
35
- function* appendText(state, text, reasoning=false) {
38
+ function* appendText(state, text, reasoning=false, indexed=false) {
36
39
  if(typeof text!=='string')throw protocolError('Interactions text delta is not a string.')
37
40
  if(!text)return
38
41
  if(state.closed)throw protocolError('Received a delta for an already closed step.')
39
42
  checkBytes(text)
40
43
  if(state.blockIndex===undefined){state.blockIndex=blockIndex++;yield{type:'block-start',index:state.blockIndex,blockType:reasoning?'reasoning':'text'}}
41
- state.text+=text;emitted=true;progressVersion++
44
+ state.text+=text;emitted=true;if(!indexed)unindexedVisibleOutput=true;progressVersion++
42
45
  yield{type:reasoning?'reasoning-delta':'text-delta',index:state.blockIndex,text}
43
46
  }
44
- function* closeText(state) {
47
+ function* closeText(state, indexed=true) {
45
48
  if(state.blockIndex!==undefined&&!state.blockClosed){
49
+ if(!indexed)unindexedVisibleOutput=true
46
50
  const block={type:state.kind==='thought'?'reasoning':'text',text:state.text}
47
51
  state.blockClosed=true;blocks.set(state.blockIndex,block)
48
52
  yield{type:'block-end',index:state.blockIndex,block}
@@ -54,14 +58,7 @@ export async function* streamInteractions(options, settings) {
54
58
  if(!state)throw protocolError('Interactions delta arrived before its step.start.')
55
59
  return state
56
60
  }
57
- function* handle(event) {
58
- if(!validObject(event))throw protocolError('Invalid Interactions event.')
59
- if(event.event_id){
60
- if(typeof event.event_id!=='string')throw protocolError('Invalid event ID.')
61
- if(seenEvents.has(event.event_id))return
62
- if(seenEvents.size>=MAX_EVENTS)throw protocolError('Interactions event retention limit reached.')
63
- seenEvents.add(event.event_id);lastEventId=event.event_id
64
- }
61
+ function* handleBody(event, indexed) {
65
62
  settings.onEvent?.(event)
66
63
  const type=event.event_type??event.type
67
64
  const id=event.interaction?.id??event.interaction_id
@@ -109,8 +106,8 @@ export async function* streamInteractions(options, settings) {
109
106
  if(!validObject(step)||!['thought','model_output','function_call'].includes(step.type))throw protocolError('Unsupported Interactions step type; this adapter executes only DSH-declared local tools.')
110
107
  const state={kind:step.type,text:'',signature:step.signature,callId:step.id,name:step.name,argsValue:step.arguments,argsText:'',closed:false}
111
108
  steps.set(event.index,state);progressVersion++
112
- if(step.type==='model_output')for(const content of step.content??[]){if(content.type!=='text')throw protocolError('Unsupported output modality.');yield*appendText(state,content.text)}
113
- if(step.type==='thought')for(const content of step.summary??[]){if(content.type==='text')yield*appendText(state,content.text,true)}
109
+ if(step.type==='model_output')for(const content of step.content??[]){if(content.type!=='text')throw protocolError('Unsupported output modality.');yield*appendText(state,content.text,false,indexed)}
110
+ if(step.type==='thought')for(const content of step.summary??[]){if(content.type==='text')yield*appendText(state,content.text,true,indexed)}
114
111
  return
115
112
  }
116
113
  if(type==='step.delta'){
@@ -121,9 +118,9 @@ export async function* streamInteractions(options, settings) {
121
118
  return
122
119
  }
123
120
  if(delta.type==='text_annotation_delta'&&state.kind==='model_output')return // additive citations; no DSH annotation block type
124
- if(delta.type==='text'&&state.kind==='model_output')yield*appendText(state,delta.text)
121
+ if(delta.type==='text'&&state.kind==='model_output')yield*appendText(state,delta.text,false,indexed)
125
122
  else if(delta.type==='thought_summary'&&state.kind==='thought'){
126
- if(delta.content?.type==='text')yield*appendText(state,delta.content.text,true)
123
+ if(delta.content?.type==='text')yield*appendText(state,delta.content.text,true,indexed)
127
124
  else if(delta.content!==undefined)throw protocolError('Unsupported thought summary modality.')
128
125
  }else if(delta.type==='thought_signature'&&state.kind==='thought'){
129
126
  if(typeof delta.signature!=='string')throw protocolError('Invalid thought signature.')
@@ -146,7 +143,7 @@ export async function* streamInteractions(options, settings) {
146
143
  if(state.kind==='function_call'){
147
144
  try{state.args=state.argsText?JSON.parse(state.argsText):(state.argsValue??{})}catch{state.invalidArguments=true}
148
145
  if(!validObject(state.args))state.invalidArguments=true
149
- }else yield*closeText(state)
146
+ }else yield*closeText(state, indexed)
150
147
  return
151
148
  }
152
149
  // Usage/status extensions may be added without rendering; unknown semantic
@@ -159,6 +156,18 @@ export async function* streamInteractions(options, settings) {
159
156
  }
160
157
  throw protocolError('Unknown Interactions stream event.')
161
158
  }
159
+ function* handle(event) {
160
+ if(!validObject(event))throw protocolError('Invalid Interactions event.')
161
+ const indexed=Boolean(event.event_id)
162
+ if(indexed){
163
+ if(typeof event.event_id!=='string')throw protocolError('Invalid event ID.')
164
+ if(seenEvents.has(event.event_id))return
165
+ if(seenEvents.size>=MAX_EVENTS)throw protocolError('Interactions event retention limit reached.')
166
+ seenEvents.add(event.event_id);lastEventId=event.event_id;unindexedVisibleOutput=false
167
+ }
168
+ yield*handleBody(event, indexed)
169
+ if(indexed)indexedSnapshot=captureSnapshot()
170
+ }
162
171
  const cancel=async()=>{
163
172
  if(settings.background&&interactionId&&!terminal.has(status)&&!cancelled){
164
173
  cancelled=true // one attempt per unwind, even if confirmation fails
@@ -169,10 +178,10 @@ export async function* streamInteractions(options, settings) {
169
178
  signal?.throwIfAborted()
170
179
  const body=await buildInteractionRequest(options,settings)
171
180
  let stream=await transport.create(body,{signal})
172
- let reconnects=0,connections=1,replayingWithoutCursor=false
181
+ let reconnects=0,connections=1,replayingWithoutCursor=false,lastDropFingerprint
173
182
  const deadline=Date.now()+24*60*60*1000
174
183
  for(;;){
175
- const beforeProgress=progressVersion
184
+ const beforeProgress=progressVersion,beforeLastEventId=lastEventId
176
185
  let failure
177
186
  try{
178
187
  if(!stream?.[Symbol.asyncIterator])throw protocolError('Interactions create/get returned no event stream.')
@@ -184,9 +193,13 @@ export async function* streamInteractions(options, settings) {
184
193
  if(!failure&&completedEvent&&terminal.has(status))break
185
194
  if(failure&&!retryable(failure))throw failure
186
195
  signal?.throwIfAborted()
187
- if((!replayingWithoutCursor||lastEventId)&&progressVersion>beforeProgress)reconnects=0
196
+ const attemptFingerprint=JSON.stringify({lastEventId,status,usage,steps:[...steps.entries()].sort((a,b)=>a[0]-b[0])})
197
+ const madeProgress=lastEventId!==beforeLastEventId||((!replayingWithoutCursor||Boolean(lastEventId))&&progressVersion>beforeProgress&&attemptFingerprint!==lastDropFingerprint)
198
+ if(madeProgress)reconnects=0
199
+ lastDropFingerprint=attemptFingerprint
188
200
  if(connections>=64||Date.now()>=deadline)throw Object.assign(new Error('Interactions recovery reached its connection or runtime bound.'),{code:'ARGON_INTERACTIONS_INTERRUPTED'})
189
201
  if(!settings.store||!interactionId||reconnects>=3)throw failure??Object.assign(new Error('Interactions stream ended without a terminal status or safe recovery cursor.'),{code:'ARGON_INTERACTIONS_INTERRUPTED'})
202
+ if(unindexedVisibleOutput)throw protocolError('The stored Interactions read stream omitted event_id on a visible output frame; refusing unsafe replay. The remote task will be explicitly terminated, not resubmitted.')
190
203
  if(!lastEventId){
191
204
  if(emitted)throw protocolError('The stored Interactions read stream omitted event_id after visible output; refusing unsafe replay. The remote task will be explicitly terminated, not resubmitted.')
192
205
  // No visible blocks or tool calls have escaped: discard cursor-free
@@ -194,7 +207,10 @@ export async function* streamInteractions(options, settings) {
194
207
  replayingWithoutCursor=true
195
208
  steps.clear();blocks.clear();seenEvents.clear();emittedToolIds.clear()
196
209
  blockIndex=0;bytes=0;status=undefined;usage=undefined;completedEvent=false
197
- }else replayingWithoutCursor=false
210
+ }else{
211
+ replayingWithoutCursor=false
212
+ if(indexedSnapshot)restoreSnapshot(indexedSnapshot)
213
+ }
198
214
  const delayMs=computeReconnectDelayMs(++reconnects,settings.reconnectBackoff,parseUpstreamRetryDelayMs(failure))
199
215
  await waitWithAbort(delayMs,signal)
200
216
  // The transport options of get() stay byte-for-byte unchanged; hops only
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "dsh-vertex-argon",
3
- "version": "2.2.6",
3
+ "version": "2.2.8",
4
4
  "description": "Gemini 4 Adapter for DSH: official Vertex AI & Gemini API channels (Argon, Carbon, Hydrogen & Neon), resilient long decode, explicit service tiers, and request diagnostics.",
5
5
  "license": "MIT",
6
6
  "repository": {