@lota-sdk/core 0.1.18 → 0.1.19

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lota-sdk/core",
-  "version": "0.1.18",
+  "version": "0.1.19",
   "type": "module",
   "main": "./src/index.ts",
   "types": "./src/index.ts",
@@ -32,7 +32,7 @@
     "@chat-adapter/slack": "^4.23.0",
     "@chat-adapter/state-ioredis": "^4.23.0",
     "@logtape/logtape": "^2.0.5",
-    "@lota-sdk/shared": "0.1.18",
+    "@lota-sdk/shared": "0.1.19",
     "@mendable/firecrawl-js": "^4.17.0",
     "@surrealdb/node": "^3.0.3",
     "ai": "^6.0.137",

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

@@ -5,6 +5,7 @@ import type { LanguageModelMiddleware } from 'ai'
 
 import { getRuntimeConfig } from '../runtime/runtime-config'
 import { isRecord, readString } from '../utils/string'
+import { buildAiGatewayCacheHeaders } from './cache-headers'
 
 type AiGatewayLanguageModel = Parameters<typeof wrapLanguageModel>[0]['model']
 type AiGatewayExtraParams = Record<string, unknown>
@@ -17,6 +18,7 @@ type AiGatewayStreamResult = Awaited<ReturnType<WrapStreamOptions['doStream']>>
 type AiGatewayGeneratedContent = AiGatewayGenerateResult['content'][number]
 type AiGatewayStreamPart = AiGatewayStreamResult['stream'] extends ReadableStream<infer T> ? T : never
 type AiGatewayConfig = { apiKey: string; baseURL: string }
+type AiGatewayProviderOptions = NonNullable<AiGatewayCallOptions['providerOptions']>
 
 const EXPECTED_GATEWAY_KEY_PREFIX = 'sk-bf-'
 const AI_GATEWAY_VIRTUAL_KEY_HEADER = 'x-bf-vk'
@@ -26,6 +28,38 @@ const OPENROUTER_RESPONSE_HEALING_EXTRA_PARAMS = {
   plugins: [{ id: 'response-healing' }],
 } as const satisfies AiGatewayExtraParams
 
+function toAiGatewayCacheKeyPart(value: string): string {
+  const normalized = value
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9:_-]+/g, '-')
+    .replace(/-+/g, '-')
+  return normalized.replace(/^-+|-+$/g, '') || 'request'
+}
+
+function mergeAiGatewayHeaders(
+  existingHeaders: AiGatewayCallOptions['headers'] | undefined,
+  additionalHeaders: Record<string, string>,
+): Record<string, string> {
+  const merged = new Headers(existingHeaders as HeadersInit | undefined)
+  for (const [key, value] of Object.entries(additionalHeaders)) {
+    if (!merged.has(key)) {
+      merged.set(key, value)
+    }
+  }
+  return Object.fromEntries(merged.entries())
+}
+
+function withDefaultAiGatewayCacheHeaders(params: AiGatewayCallOptions, modelId: string): AiGatewayCallOptions {
+  return {
+    ...params,
+    headers: mergeAiGatewayHeaders(
+      params.headers,
+      buildAiGatewayCacheHeaders(`model:${toAiGatewayCacheKeyPart(modelId)}`),
+    ),
+  }
+}
+
 function normalizeAiGatewayUrl(value: string): string {
   const trimmed = value.trim()
   if (!trimmed) {
@@ -37,10 +71,10 @@ function normalizeAiGatewayUrl(value: string): string {
 }
 
 function readDirectEnvAiGatewayConfig(): AiGatewayConfig {
-  const apiKey = (process.env.LOTA_KEY ?? process.env.VENTUROS_KEY ?? '').trim()
+  const apiKey = (process.env.AI_GATEWAY_KEY ?? '').trim()
   if (!apiKey) {
     throw new Error(
-      '[ai-gateway] Missing AI gateway key. Set LOTA_KEY or VENTUROS_KEY, or configure createLotaRuntime({ aiGateway: { key } }).',
+      '[ai-gateway] Missing AI gateway key. Set AI_GATEWAY_KEY, or configure createLotaRuntime({ aiGateway: { key } }).',
     )
   }
 
@@ -234,6 +268,30 @@ function addAiGatewayReasoningRawChunks(
   return { ...params, includeRawChunks: true }
 }
 
+export function normalizeAiGatewayChatProviderOptions(params: AiGatewayCallOptions): AiGatewayCallOptions {
+  const providerOptions = isRecord(params.providerOptions)
+    ? ({ ...params.providerOptions } as AiGatewayProviderOptions)
+    : ({} as AiGatewayProviderOptions)
+  const openaiOptions = isRecord(providerOptions.openai)
+    ? { ...providerOptions.openai }
+    : ({} as Record<string, unknown>)
+
+  if (openaiOptions.systemMessageMode === 'system') {
+    return params
+  }
+
+  return {
+    ...params,
+    providerOptions: {
+      ...providerOptions,
+      openai: {
+        ...openaiOptions,
+        ...(openaiOptions.systemMessageMode === 'remove' ? {} : { systemMessageMode: 'system' }),
+      },
+    },
+  }
+}
+
 export function injectAiGatewayExtraParamsRequestBody(
   body: BodyInit | null | undefined,
   extraParams: AiGatewayExtraParams,
@@ -309,7 +367,8 @@ export function aiGatewayModel(modelId: string) {
       model: getAiGatewayProvider()(modelId),
       middleware: {
         specificationVersion: 'v3',
-        transformParams: async ({ params, type }) => addAiGatewayReasoningRawChunks(params, type),
+        transformParams: async ({ params, type }) =>
+          withDefaultAiGatewayCacheHeaders(addAiGatewayReasoningRawChunks(params, type), modelId),
         wrapStream: async ({ doStream, params }) => {
           const result = await doStream()
           if (!isReasoningEnabled(params)) return result
@@ -322,7 +381,15 @@ export function aiGatewayModel(modelId: string) {
 }
 
 export function aiGatewayOpenRouterResponseHealingModel(modelId: string) {
-  return withAiGatewayDevTools(getAiGatewayOpenRouterResponseHealingProvider()(modelId))
+  return withAiGatewayDevTools(
+    wrapLanguageModel({
+      model: getAiGatewayOpenRouterResponseHealingProvider()(modelId),
+      middleware: {
+        specificationVersion: 'v3',
+        transformParams: async ({ params }) => withDefaultAiGatewayCacheHeaders(params, modelId),
+      },
+    }),
+  )
 }
 
 export function aiGatewayChatModel(modelId: string) {
@@ -331,7 +398,10 @@ export function aiGatewayChatModel(modelId: string) {
       model: getAiGatewayProvider().chat(modelId),
       middleware: {
         specificationVersion: 'v3',
-        transformParams: async ({ params, type }) => addAiGatewayReasoningRawChunks(params, type),
+        transformParams: async ({ params, type }) =>
+          normalizeAiGatewayChatProviderOptions(
+            withDefaultAiGatewayCacheHeaders(addAiGatewayReasoningRawChunks(params, type), modelId),
+          ),
         wrapGenerate: async ({ doGenerate }) => {
           const result = await doGenerate()
 

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

@@ -1,8 +1,33 @@
 const AI_GATEWAY_CACHE_KEY_HEADER = 'x-bf-cache-key'
 const AI_GATEWAY_CACHE_TTL_HEADER = 'x-bf-cache-ttl'
+const AI_GATEWAY_CACHE_THRESHOLD_HEADER = 'x-bf-cache-threshold'
+const AI_GATEWAY_CACHE_TYPE_HEADER = 'x-bf-cache-type'
 
-export function buildAiGatewayCacheHeaders(cacheKey: string, ttl?: string): Record<string, string> {
+export const AI_GATEWAY_STRICT_SEMANTIC_CACHE_THRESHOLD = 0.975
+
+export type AiGatewayCacheType = 'direct' | 'semantic'
+
+export function buildAiGatewayCacheHeaders(
+  cacheKey: string,
+  ttl?: string,
+  threshold?: number,
+  cacheType?: AiGatewayCacheType,
+): Record<string, string> {
   const headers: Record<string, string> = { [AI_GATEWAY_CACHE_KEY_HEADER]: cacheKey }
   if (ttl) headers[AI_GATEWAY_CACHE_TTL_HEADER] = ttl
+  if (typeof threshold === 'number') headers[AI_GATEWAY_CACHE_THRESHOLD_HEADER] = String(threshold)
+  if (cacheType) headers[AI_GATEWAY_CACHE_TYPE_HEADER] = cacheType
   return headers
 }
+
+export function buildAiGatewayDirectCacheHeaders(cacheKey: string, ttl?: string): Record<string, string> {
+  return buildAiGatewayCacheHeaders(cacheKey, ttl, undefined, 'direct')
+}
+
+export function buildAiGatewayStrictSemanticCacheHeaders(
+  cacheKey: string,
+  ttl?: string,
+  threshold = AI_GATEWAY_STRICT_SEMANTIC_CACHE_THRESHOLD,
+): Record<string, string> {
+  return buildAiGatewayCacheHeaders(cacheKey, ttl, threshold, 'semantic')
+}

package/src/runtime/chat-run-registry.ts

@@ -1,6 +1,10 @@
 export class ChatRunRegistry {
   private controllers = new Map<string, AbortController>()
 
+  has(runId: string): boolean {
+    return this.controllers.has(runId)
+  }
+
   register(runId: string, controller: AbortController): void {
     this.controllers.set(runId, controller)
   }

package/src/runtime/context-compaction.ts

@@ -101,6 +101,90 @@ function sanitizeStateText(value: string): string | null {
   return normalized
 }
 
+function buildExistingWorkstreamStateForCompactionPrompt(state: WorkstreamState): Record<string, unknown> {
+  const currentPlanText = state.currentPlan ? sanitizeStateText(state.currentPlan.text) : null
+
+  const activeConstraints = state.activeConstraints
+    .map((constraint) => ({
+      id: constraint.id,
+      text: sanitizeStateText(constraint.text),
+      source: constraint.source,
+      approved: constraint.approved,
+      sourceMessageIds: constraint.sourceMessageIds,
+    }))
+    .filter((constraint): constraint is typeof constraint & { text: string } => Boolean(constraint.text))
+
+  const keyDecisions = state.keyDecisions
+    .map((decision) => ({
+      id: decision.id,
+      decision: sanitizeStateText(decision.decision),
+      rationale: sanitizeStateText(decision.rationale),
+      agent: decision.agent,
+      sourceMessageIds: decision.sourceMessageIds,
+      confidence: decision.confidence,
+    }))
+    .filter((decision): decision is typeof decision & { decision: string; rationale: string } =>
+      Boolean(decision.decision && decision.rationale),
+    )
+
+  const tasks = state.tasks
+    .map((task) => ({
+      id: task.id,
+      title: sanitizeStateText(task.title),
+      status: task.status,
+      owner: sanitizeStateText(task.owner),
+      externalId: task.externalId,
+      source: task.source,
+      sourceMessageIds: task.sourceMessageIds,
+    }))
+    .filter((task): task is typeof task & { title: string; owner: string } => Boolean(task.title && task.owner))
+
+  const openQuestions = state.openQuestions
+    .map((question) => ({
+      id: question.id,
+      text: sanitizeStateText(question.text),
+      source: question.source,
+      sourceMessageIds: question.sourceMessageIds,
+    }))
+    .filter((question): question is typeof question & { text: string } => Boolean(question.text))
+
+  const artifacts = state.artifacts
+    .map((artifact) => ({
+      id: artifact.id,
+      name: sanitizeStateText(artifact.name),
+      type: artifact.type,
+      pointer: sanitizeStateText(artifact.pointer),
+    }))
+    .filter((artifact): artifact is typeof artifact & { name: string; pointer: string } =>
+      Boolean(artifact.name && artifact.pointer),
+    )
+
+  const agentContributions = state.agentContributions
+    .map((note) => ({ id: note.id, agent: note.agent, summary: sanitizeStateText(note.summary) }))
+    .filter((note): note is typeof note & { summary: string } => Boolean(note.summary))
+
+  return {
+    currentPlan: currentPlanText
+      ? {
+          id: state.currentPlan?.id,
+          text: currentPlanText,
+          source: state.currentPlan?.source,
+          approved: state.currentPlan?.approved,
+          sourceMessageIds: state.currentPlan?.sourceMessageIds ?? [],
+        }
+      : null,
+    activeConstraints,
+    keyDecisions,
+    tasks,
+    openQuestions,
+    risks: state.risks.map((risk) => sanitizeStateText(risk)).filter((risk): risk is string => Boolean(risk)),
+    artifacts,
+    agentContributions,
+    approvedBy: state.approvedBy ? sanitizeStateText(state.approvedBy) : undefined,
+    approvalNote: state.approvalNote ? sanitizeStateText(state.approvalNote) : undefined,
+  }
+}
+
 function createStableId(prefix: string, ...parts: Array<string | number | undefined>): string {
   const payload = parts
     .map((part) => (part === undefined ? '' : String(part)))
@@ -488,7 +572,7 @@ export function buildContextCompactionPrompt(params: ContextCompactionPromptPara
     params.previousSummary.trim() || 'None',
     '</previous-summary>',
     '<existing-workstream-state>',
-    JSON.stringify(params.existingState),
+    JSON.stringify(buildExistingWorkstreamStateForCompactionPrompt(params.existingState)),
     '</existing-workstream-state>',
     '<new-messages>',
     params.transcript || 'None',
@@ -605,27 +689,34 @@ export function createContextCompactionRuntime(
       .filter((constraint): constraint is typeof constraint & { text: string } => Boolean(constraint.text))
 
     const openQuestions = state.openQuestions
-      .map((question) => ({ ...question, text: sanitizeStateText(question.text) }))
-      .filter((question): question is typeof question & { text: string } => Boolean(question.text))
+      .map((question) => sanitizeStateText(question.text))
+      .filter((question): question is string => Boolean(question))
 
     const decisions = state.keyDecisions
       .map((decision) => ({
-        ...decision,
+        agent: decision.agent,
         decision: sanitizeStateText(decision.decision),
         rationale: sanitizeStateText(decision.rationale),
+        confidence: decision.confidence,
       }))
       .filter((decision): decision is typeof decision & { decision: string; rationale: string } =>
         Boolean(decision.decision && decision.rationale),
       )
 
     const tasks = state.tasks
-      .map((task) => ({ ...task, title: sanitizeStateText(task.title), owner: sanitizeStateText(task.owner) }))
+      .map((task) => ({
+        title: sanitizeStateText(task.title),
+        status: task.status,
+        owner: sanitizeStateText(task.owner),
+        externalId: task.externalId,
+        source: task.source,
+      }))
       .filter((task): task is typeof task & { title: string; owner: string } => Boolean(task.title && task.owner))
 
     const artifacts = state.artifacts
       .map((artifact) => ({
-        ...artifact,
         name: sanitizeStateText(artifact.name),
+        type: artifact.type,
         pointer: sanitizeStateText(artifact.pointer),
       }))
       .filter((artifact): artifact is typeof artifact & { name: string; pointer: string } =>
@@ -633,7 +724,7 @@ export function createContextCompactionRuntime(
       )
 
     const agentContributions = state.agentContributions
-      .map((note) => ({ ...note, summary: sanitizeStateText(note.summary) }))
+      .map((note) => ({ agent: note.agent, summary: sanitizeStateText(note.summary) }))
       .filter((note): note is typeof note & { summary: string } => Boolean(note.summary))
 
     const payload = {
@@ -661,12 +752,9 @@ export function createContextCompactionRuntime(
       artifacts,
       agentContributions,
       advisory: {
-        approvedBy: state.approvedBy ?? null,
-        approvedAt: state.approvedAt ?? null,
-        approvalMessageId: state.approvalMessageId ?? null,
-        approvalNote: state.approvalNote ?? null,
+        approvedBy: state.approvedBy ? sanitizeStateText(state.approvedBy) : null,
+        approvalNote: state.approvalNote ? sanitizeStateText(state.approvalNote) : null,
       },
-      lastUpdated: state.lastUpdated,
     }
 
     return ['<workstream-state>', JSON.stringify(payload, null, 2), '</workstream-state>'].join('\n')

package/src/runtime/memory-prompts-fact.ts

@@ -1,3 +1,5 @@
+import { formatUtcPromptDate } from '../utils/date-time'
+
 export function getFactRetrievalMessages(
   parsedMessages: string,
   customPrompt?: string,
@@ -38,7 +40,7 @@ Hard rules:
 - Prefer returning fewer items. If uncertain, return an empty list.
 - Max ${maxFacts} facts.
 
-Today's date is ${new Date().toISOString().split('T')[0]}.
+Today's date is ${formatUtcPromptDate(new Date())}.
 ${baseInstructions}`
 
   const userPrompt = `Conversation:\n${parsedMessages}`

package/src/runtime/runtime-config.ts

@@ -292,7 +292,7 @@ export const LOTA_RUNTIME_ENV_KEYS = Object.freeze([
   'SURREALDB_PASSWORD',
   'REDIS_URL',
   'AI_GATEWAY_URL',
-  'LOTA_KEY',
+  'AI_GATEWAY_KEY',
   'AI_EMBEDDING_MODEL',
   'S3_ENDPOINT',
   'S3_BUCKET',

package/src/services/workstream-title.service.ts

@@ -10,7 +10,7 @@ import {
 } from '../system-agents/title-generator.agent'
 import { workstreamService } from './workstream.service'
 
-const WORKSTREAM_TITLE_TIMEOUT_MS = 5_000
+const WORKSTREAM_TITLE_TIMEOUT_MS = 30_000
 
 class WorkstreamTitleService {
   helperRuntime = createHelperModelRuntime()

package/src/services/workstream-turn-preparation.service.ts

@@ -632,7 +632,10 @@ export async function prepareWorkstreamRunCore(params: WorkstreamRunCoreParams):
   const workstreamRecord = await waitForWorkstreamCompactionIfNeeded(workstreamRef)
   timer.step('compaction-gate')
   if (toOptionalTrimmedString(workstreamRecord.activeRunId)) {
-    throw new WorkstreamTurnError('A chat run is already active.', 409)
+    const clearedStaleRun = await workstreamService.clearStaleActiveRunIfMissingFromRegistry(workstreamRef)
+    if (!clearedStaleRun) {
+      throw new WorkstreamTurnError('A chat run is already active.', 409)
+    }
   }
 
   if (params.kind === 'approvalContinuation' || params.kind === 'nativeToolApprovalTurn') {

package/src/services/workstream.service.ts

@@ -525,6 +525,22 @@ class WorkstreamService extends BaseService<typeof WorkstreamSchema> {
     )
   }
 
+  async clearStaleActiveRunIfMissingFromRegistry(workstreamId: RecordIdRef): Promise<boolean> {
+    const activeRunId = await this.getActiveRunId(workstreamId)
+    if (!activeRunId || chatRunRegistry.has(activeRunId)) {
+      return false
+    }
+
+    const activeStreamId = await this.getActiveStreamId(workstreamId)
+    await Promise.all([
+      this.clearActiveRunIdIfMatches(workstreamId, activeRunId),
+      activeStreamId ? this.clearActiveStreamIdIfMatches(workstreamId, activeStreamId) : Promise.resolve(),
+    ])
+
+    serverLogger.warn`Cleared stale workstream run after process restart: workstream=${recordIdToString(ensureRecordId(workstreamId, TABLES.WORKSTREAM), TABLES.WORKSTREAM)} run=${activeRunId}`
+    return true
+  }
+
   async stopActiveRun(workstreamId: RecordIdRef): Promise<boolean> {
     const activeRunId = await this.getActiveRunId(workstreamId)
     if (!activeRunId) return false
@@ -534,7 +550,7 @@ class WorkstreamService extends BaseService<typeof WorkstreamSchema> {
       return true
     }
 
-    await this.clearActiveRunIdIfMatches(workstreamId, activeRunId)
+    await this.clearStaleActiveRunIfMissingFromRegistry(workstreamId)
     return false
   }
 
@@ -675,6 +691,7 @@ class WorkstreamService extends BaseService<typeof WorkstreamSchema> {
       typeof workstream.activeRunId === 'string' && workstream.activeRunId.trim().length > 0
         ? workstream.activeRunId
         : null
+    const isRunning = activeRunId !== null && chatRunRegistry.has(activeRunId)
     const isCompacting = workstream.isCompacting === true
     const mode = workstream.mode
     const core = workstream.core
@@ -688,7 +705,7 @@ class WorkstreamService extends BaseService<typeof WorkstreamSchema> {
       core,
       ...(coreType ? { coreType } : {}),
       nameGenerated: workstream.nameGenerated,
-      isRunning: activeRunId !== null,
+      isRunning,
       isCompacting,
       ...(isAgentName(workstream.agentId) ? { agentId: workstream.agentId } : {}),
       title: workstream.title ?? this.getDefaultTitle(workstream),

package/src/system-agents/context-compaction.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -41,7 +41,7 @@ export function createContextCompactionAgent(options: CreateHelperToolLoopAgentO
   return new ToolLoopAgent({
     id: 'context-compaction',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('context-compaction'),
+    headers: buildAiGatewayDirectCacheHeaders('context-compaction'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, { instructions: CONTEXT_COMPACTION_PROMPT }),
   })

package/src/system-agents/delegated-agent-factory.ts

@@ -4,6 +4,7 @@ import { z } from 'zod'
 
 import type { ToolDefinition } from '../ai/definitions'
 import { aiLogger } from '../config/logger'
+import { formatUtcPromptDate } from '../utils/date-time'
 import { isRecord } from '../utils/string'
 import { assertSubstantiveAgentResult } from './agent-result'
 
@@ -32,15 +33,7 @@ function resolveAgentModel(model: AgentModel): LanguageModel {
 }
 
 function buildCurrentDateContext(now = new Date()): string {
-  const isoDate = now.toISOString().slice(0, 10)
-  const humanDate = new Intl.DateTimeFormat('en-US', {
-    timeZone: 'UTC',
-    year: 'numeric',
-    month: 'long',
-    day: 'numeric',
-  }).format(now)
-
-  return [`Today is ${isoDate} (${humanDate}, UTC).`, 'Use this exact date for any recency reasoning.'].join(' ')
+  return [`Today is ${formatUtcPromptDate(now)}.`, 'Use this exact date for any recency reasoning.'].join(' ')
 }
 
 export function buildRecencyInstructions(tools?: ToolSet): string {

package/src/system-agents/memory-reranker.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
@@ -33,7 +33,7 @@ export function createMemoryRerankerAgent(options: CreateHelperToolLoopAgentOpti
   return new ToolLoopAgent({
     id: 'memory-reranker',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('memory-reranker'),
+    headers: buildAiGatewayDirectCacheHeaders('memory-reranker'),
     providerOptions: OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options),
   })

package/src/system-agents/memory.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -53,7 +53,7 @@ export function createOrgMemoryAgent(options: CreateHelperToolLoopAgentOptions)
   return new ToolLoopAgent({
     id: 'org-memory',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('org-memory'),
+    headers: buildAiGatewayDirectCacheHeaders('org-memory'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options),
   })

package/src/system-agents/recent-activity-title-refiner.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import { getLeadAgentDisplayName } from '../config/agent-defaults'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
@@ -80,7 +80,7 @@ export function createRecentActivityTitleRefinerAgent(options: CreateHelperToolL
   return new ToolLoopAgent({
     id: 'recent-activity-title-refiner',
     model: aiGatewayModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('recent-activity-title-refiner'),
+    headers: buildAiGatewayDirectCacheHeaders('recent-activity-title-refiner'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, {
       instructions: buildRecentActivityTitleRefinerPrompt(),

package/src/system-agents/regular-chat-memory-digest.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -67,7 +67,7 @@ export function createRegularChatMemoryDigestAgent(options: CreateHelperToolLoop
   return new ToolLoopAgent({
     id: 'regular-chat-memory-digest',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('regular-chat-memory-digest'),
+    headers: buildAiGatewayDirectCacheHeaders('regular-chat-memory-digest'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, {
       instructions: regularChatMemoryDigestPrompt,

package/src/system-agents/skill-extractor.agent.ts

@@ -2,7 +2,7 @@ import { ToolLoopAgent } from 'ai'
 import { z } from 'zod'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -80,7 +80,7 @@ export function createSkillExtractorAgent(options: CreateHelperToolLoopAgentOpti
   return new ToolLoopAgent({
     id: 'skill-extractor',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('skill-extractor'),
+    headers: buildAiGatewayDirectCacheHeaders('skill-extractor'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, {
       instructions: skillExtractorPrompt,

package/src/system-agents/skill-manager.agent.ts

@@ -2,7 +2,7 @@ import { ToolLoopAgent } from 'ai'
 import { z } from 'zod'
 
 import { aiGatewayOpenRouterResponseHealingModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -70,7 +70,7 @@ export function createSkillManagerAgent(options: CreateHelperToolLoopAgentOption
   return new ToolLoopAgent({
     id: 'skill-manager',
     model: aiGatewayOpenRouterResponseHealingModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('skill-manager'),
+    headers: buildAiGatewayDirectCacheHeaders('skill-manager'),
     providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, {
       instructions: skillManagerPrompt,

package/src/system-agents/title-generator.agent.ts

@@ -1,7 +1,7 @@
 import { ToolLoopAgent } from 'ai'
 
 import { aiGatewayModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_FAST_REASONING_MODEL_ID,
   OPENROUTER_MINIMAL_REASONING_PROVIDER_OPTIONS,
@@ -34,7 +34,7 @@ export function createWorkstreamTitleGeneratorAgent(options: CreateHelperToolLoo
   return new ToolLoopAgent({
     id: 'workstream-title-generator',
     model: aiGatewayModel(OPENROUTER_FAST_REASONING_MODEL_ID),
-    headers: buildAiGatewayCacheHeaders('workstream-title-generator'),
+    headers: buildAiGatewayDirectCacheHeaders('workstream-title-generator'),
     providerOptions: OPENROUTER_MINIMAL_REASONING_PROVIDER_OPTIONS,
     ...resolveHelperAgentOptions(options, {
       instructions: WORKSTREAM_TITLE_GENERATOR_PROMPT,

package/src/tools/research-topic.tool.ts

@@ -1,5 +1,5 @@
 import { aiGatewayChatModel } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayCacheHeaders } from '../ai-gateway/cache-headers'
+import { buildAiGatewayStrictSemanticCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_MEDIUM_REASONING_PROVIDER_OPTIONS,
   OPENROUTER_WEB_RESEARCH_MODEL_ID,
@@ -15,7 +15,7 @@ export const researchTopicTool = createDelegatedAgentTool({
     'Delegate a research task to a dedicated research agent that searches the web, fetches pages, and returns a synthesized markdown report. Call multiple instances in parallel for broad research across different topics.',
   model: () => aiGatewayChatModel(OPENROUTER_WEB_RESEARCH_MODEL_ID),
   providerOptions: OPENROUTER_MEDIUM_REASONING_PROVIDER_OPTIONS,
-  headers: buildAiGatewayCacheHeaders('researchTopic'),
+  headers: buildAiGatewayStrictSemanticCacheHeaders('researchTopic'),
   instructions: RESEARCHER_PROMPT,
   tools: { searchWeb: searchWebTool.create(), fetchWebpage: fetchWebpageTool.create() },
 })

package/src/utils/date-time.ts

@@ -1,6 +1,17 @@
 export { toIsoDateTimeString, toOptionalIsoDateTimeString } from '@lota-sdk/shared'
 
+const PROMPT_DATE_FORMATTER = new Intl.DateTimeFormat('en-US', {
+  timeZone: 'UTC',
+  year: 'numeric',
+  month: 'long',
+  day: 'numeric',
+})
+
 export function toDatabaseDateTime(value: string | Date | null | undefined): Date | undefined {
   if (value === null || value === undefined) return undefined
   return value instanceof Date ? value : new Date(value)
 }
+
+export function formatUtcPromptDate(value: Date): string {
+  return PROMPT_DATE_FORMATTER.format(value)
+}

package/src/workers/utils/file-section-chunker.ts

@@ -1,6 +1,6 @@
 import { CHARS_PER_TOKEN_ESTIMATE } from '../../utils/string'
 
-export const DEFAULT_FILE_SECTION_CHUNK_MAX_CHARS = 250_000
+export const DEFAULT_FILE_SECTION_CHUNK_MAX_CHARS = 1_500_000
 export const MIN_FILE_SECTION_CHUNK_MAX_CHARS = 4_000
 export const DEFAULT_FILE_SECTION_CHUNK_MIN_CHARS = 10_000
 const SECTION_SEPARATOR_LENGTH = 2