@lota-sdk/core 0.4.45 → 0.4.47

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lota-sdk/core",
-  "version": "0.4.45",
+  "version": "0.4.47",
   "files": [
     "src",
     "infrastructure/schema"
@@ -32,7 +32,7 @@
     "@ai-sdk/provider": "^3.0.9",
     "@chat-adapter/slack": "^4.26.0",
     "@chat-adapter/state-ioredis": "^4.26.0",
-    "@lota-sdk/shared": "0.4.45",
+    "@lota-sdk/shared": "0.4.47",
     "@mendable/firecrawl-js": "^4.20.0",
     "@surrealdb/node": "^3.0.3",
     "ai": "^6.0.170",

package/src/ai-gateway/ai-gateway.ts

@@ -37,6 +37,8 @@ class AiGatewayStreamAttemptTag extends Context.Service<
 
 const EXPECTED_GATEWAY_KEY_PREFIX = 'sk-bf-'
 const AI_GATEWAY_VIRTUAL_KEY_HEADER = 'x-bf-vk'
+const AI_GATEWAY_PASSTHROUGH_EXTRA_PARAMS_HEADER = 'x-bf-passthrough-extra-params'
+const AZURE_OPENAI_PROMPT_CACHE_RETENTION = '24h'
 const AI_GATEWAY_TIMEOUT_MS = 360_000
 const AI_GATEWAY_STREAM_IDLE_TIMEOUT_MS = 180_000
 const AI_GATEWAY_MAX_RETRIES = 4
@@ -512,7 +514,11 @@ export function makeAiGatewayService(
       })
     }
     const baseURL = yield* normalizeAiGatewayUrlEffect(config.aiGateway.url)
-    const provider = createOpenAI({ baseURL, apiKey, headers: { [AI_GATEWAY_VIRTUAL_KEY_HEADER]: apiKey } })
+    const provider = createOpenAI({
+      baseURL,
+      apiKey,
+      headers: { [AI_GATEWAY_VIRTUAL_KEY_HEADER]: apiKey, [AI_GATEWAY_PASSTHROUGH_EXTRA_PARAMS_HEADER]: 'true' },
+    })
 
     return AiGatewayTag.of({ semaphore, provider })
   })
@@ -728,6 +734,47 @@ function isOpenRouterModel(modelId: string): boolean {
   return modelId.trim().toLowerCase().startsWith('openrouter/')
 }
 
+function isAzureOpenAiPromptCacheModel(modelId: string): boolean {
+  const normalized = modelId.trim().toLowerCase()
+  if (!normalized) return false
+
+  const [providerPrefix, ...modelParts] = normalized.split('/')
+  if (modelParts.length > 0) {
+    if (providerPrefix !== 'azure') return false
+    return isAzureOpenAiPromptCacheModel(modelParts.join('/'))
+  }
+
+  return (
+    normalized === 'main-gpt-model' ||
+    normalized === 'mini-gpt-model' ||
+    normalized === 'nano-gpt-model' ||
+    normalized === 'gpt-5.5' ||
+    normalized.startsWith('gpt-5.5-') ||
+    normalized === 'gpt-5.4' ||
+    normalized.startsWith('gpt-5.4-')
+  )
+}
+
+export function addAzureOpenAiPromptCacheRetention(
+  params: AiGatewayCallOptions,
+  modelId?: string,
+): AiGatewayCallOptions {
+  if (!modelId || !isAzureOpenAiPromptCacheModel(modelId)) {
+    return params
+  }
+
+  const providerOptions = isRecord(params.providerOptions) ? { ...params.providerOptions } : {}
+  const openaiOptions = isRecord(providerOptions.openai) ? { ...providerOptions.openai } : {}
+
+  return {
+    ...params,
+    providerOptions: {
+      ...providerOptions,
+      openai: { ...openaiOptions, promptCacheRetention: AZURE_OPENAI_PROMPT_CACHE_RETENTION },
+    } as AiGatewayCallOptions['providerOptions'],
+  }
+}
+
 function mergeAbortSignals(signals: Array<AbortSignal | undefined>): { signal?: AbortSignal; cleanup: () => void } {
   const activeSignals = signals.filter((signal): signal is AbortSignal => Boolean(signal))
   if (activeSignals.length === 0) return { cleanup: () => undefined }
@@ -1106,7 +1153,9 @@ function createAiGatewayLanguageModelMiddleware(
       Promise.resolve(
         addAiGatewayReasoningRawChunks(
           normalizeAiGatewayJsonSchemas(
-            providerId === OPENAI_CHAT_PROVIDER_ID ? normalizeAiGatewayChatProviderOptions(params, modelId) : params,
+            providerId === OPENAI_CHAT_PROVIDER_ID
+              ? normalizeAiGatewayChatProviderOptions(addAzureOpenAiPromptCacheRetention(params, modelId), modelId)
+              : addAzureOpenAiPromptCacheRetention(params, modelId),
           ),
           type,
         ),

package/src/ai-gateway/index.ts

@@ -9,6 +9,7 @@ export {
   aiGatewayModel,
   aiGatewayOpenRouterResponseHealingModel,
   createAiGatewayModels,
+  addAzureOpenAiPromptCacheRetention,
   extractAiGatewayChatReasoningDeltaText,
   extractAiGatewayChatReasoningText,
   injectAiGatewayChatReasoningContent,

package/src/runtime/context-compaction/context-compaction.ts

@@ -11,6 +11,7 @@ import {
   readString,
   stringifyUnknown,
 } from '../../utils/string'
+import { trimOversizedTextForRequestBudget } from '../model-input-budget'
 import {
   COMPACTION_CHUNK_MAX_CHARS,
   CONTEXT_COMPACTION_INCLUDED_TOOL_NAMES,
@@ -71,6 +72,8 @@ class CompactionParseError extends Schema.TaggedErrorClass<CompactionParseError>
 }) {}
 
 const COMPACTION_RUNNER_RETRY_OPTIONS = { times: 2, schedule: Schedule.exponential(Duration.millis(500), 2) } as const
+const COMPACTION_CONTEXT_MESSAGE_TEXT_MAX_CHARS = 24_000
+const COMPACTION_MIN_SOURCE_CHARS = 2_000
 
 export interface ContextCompactionPromptParams {
   previousSummary: string
@@ -294,11 +297,14 @@ export function createContextCompactionRuntime(
     return chunks.join('\n').trim()
   }
 
-  const toContextMessageFromChatMessage = (message: ChatMessage): ContextMessage => ({
-    role: message.role,
-    text: toContextTextFromChatMessage(message),
-    sourceMessageId: message.id,
-  })
+  const toContextMessageFromChatMessage = (message: ChatMessage): ContextMessage => {
+    const text = toContextTextFromChatMessage(message)
+    return {
+      role: message.role,
+      text: trimOversizedTextForRequestBudget(text, COMPACTION_CONTEXT_MESSAGE_TEXT_MAX_CHARS),
+      sourceMessageId: message.id,
+    }
+  }
 
   const createSummaryMessage = (summaryText: string): ChatMessage | null => {
     const summary = normalizeSummary(summaryText)
@@ -458,6 +464,10 @@ export function createContextCompactionRuntime(
                 return { ...state, estimatedTokens: assessment.estimatedTokens, done: true }
               }
 
+              if (sourceText.length < COMPACTION_MIN_SOURCE_CHARS) {
+                return { ...state, estimatedTokens: assessment.estimatedTokens, done: true }
+              }
+
               const compacted = yield* compactContextMessagesEffect({
                 previousSummary: state.summaryText,
                 newMessages: contextMessages,
@@ -465,9 +475,7 @@ export function createContextCompactionRuntime(
               const rolledSummary = yield* rollupSummaryIfOversizedEffect(normalizeSummary(compacted.summary))
 
               if (rolledSummary.length >= sourceText.length) {
-                return yield* new CompactionError({
-                  message: 'Compaction summary is not shorter than compacted source',
-                })
+                return { ...state, estimatedTokens: assessment.estimatedTokens, done: true }
               }
 
               return yield* stepHistory({