@lota-sdk/core 0.4.41 → 0.4.43

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lota-sdk/core",
-  "version": "0.4.41",
+  "version": "0.4.43",
   "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": "0.4.41",
+    "@lota-sdk/shared": "0.4.43",
     "@mendable/firecrawl-js": "^4.20.0",
     "@surrealdb/node": "^3.0.3",
     "ai": "^6.0.170",

package/src/ai/index.ts

@@ -1,2 +1 @@
 export * from './definitions'
-export * from './embedding-cache'

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

@@ -10,7 +10,6 @@ import { ERROR_TAGS, AiGenerationError, ConfigurationError } from '../effect/err
 import { RuntimeConfigServiceTag } from '../effect/services'
 import { openRouterEmbeddingModel } from '../embeddings/openrouter'
 import { isRecord, readString } from '../utils/string'
-import { buildAiGatewayCacheHeaders } from './cache-headers'
 
 type AiGatewayChatResponse = { body?: unknown }
 type AiGatewayTransformParamsOptions = Parameters<NonNullable<LanguageModelMiddleware['transformParams']>>[0]
@@ -473,23 +472,6 @@ function withAiGatewayStreamIdleTimeout(
   })
 }
 
-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): AiGatewayCallOptions {
-  return { ...params, headers: mergeAiGatewayHeaders(params.headers, buildAiGatewayCacheHeaders('lota-sdk')) }
-}
-
 function normalizeAiGatewayUrlEffect(value: string): Effect.Effect<string, ConfigurationError> {
   const trimmed = value.trim()
   if (!trimmed) {
@@ -1122,13 +1104,11 @@ function createAiGatewayLanguageModelMiddleware(
     specificationVersion: 'v3',
     transformParams: ({ params, type }) =>
       Promise.resolve(
-        withDefaultAiGatewayCacheHeaders(
-          addAiGatewayReasoningRawChunks(
-            normalizeAiGatewayJsonSchemas(
-              providerId === OPENAI_CHAT_PROVIDER_ID ? normalizeAiGatewayChatProviderOptions(params, modelId) : params,
-            ),
-            type,
+        addAiGatewayReasoningRawChunks(
+          normalizeAiGatewayJsonSchemas(
+            providerId === OPENAI_CHAT_PROVIDER_ID ? normalizeAiGatewayChatProviderOptions(params, modelId) : params,
           ),
+          type,
         ),
       ),
     wrapGenerate: ({ params }) => {

package/src/ai-gateway/index.ts

@@ -20,4 +20,3 @@ export {
   normalizeAiGatewayUrl,
 } from './ai-gateway'
 export type { AiGatewayDeps, AiGatewayModels, RuntimeBridge } from './ai-gateway'
-export * from './cache-headers'

package/src/config/model-constants.ts

@@ -4,7 +4,6 @@ export {
   OPENROUTER_GEMINI_PRO_MODEL_ID,
   OPENAI_HIGH_REASONING_PROVIDER_OPTIONS,
   OPENAI_REASONING_MODEL_ID,
-  OPENROUTER_FAST_RERANK_MODEL_ID,
   OPENROUTER_FAST_REASONING_MODEL_ID,
   OPENROUTER_GEMINI_FLASH_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,

package/src/effect/errors.ts

@@ -32,7 +32,6 @@ export const ERROR_TAGS = {
   CursorPaginationError: '@lota-sdk/core/CursorPaginationError',
   DatabaseError: '@lota-sdk/core/DatabaseError',
   DelegatedAgentError: '@lota-sdk/core/DelegatedAgentError',
-  EmbeddingCacheError: '@lota-sdk/core/EmbeddingCacheError',
   EmbeddingProviderError: '@lota-sdk/core/EmbeddingProviderError',
   ExecutionPlanCacheError: '@lota-sdk/core/ExecutionPlanCacheError',
   FeedbackLoopServiceError: '@lota-sdk/core/FeedbackLoopServiceError',
@@ -73,6 +72,7 @@ export const ERROR_TAGS = {
   QueueWorkerError: '@lota-sdk/core/QueueWorkerError',
   ReadFilePartsError: '@lota-sdk/core/ReadFilePartsError',
   RedisError: '@lota-sdk/core/RedisError',
+  RedisJsonCacheError: '@lota-sdk/core/RedisJsonCacheError',
   RerankServiceError: '@lota-sdk/core/RerankServiceError',
   RuntimeLifecycleError: '@lota-sdk/core/RuntimeLifecycleError',
   SandboxedWorkerBootstrapError: '@lota-sdk/core/SandboxedWorkerBootstrapError',

package/src/embeddings/provider.ts

@@ -7,15 +7,18 @@ import { openRouterEmbeddingModel } from './openrouter'
 const SUPPORTED_EMBEDDING_PREFIXES = ['openai/', 'openrouter/'] as const
 const SUPPORTED_BARE_EMBEDDING_MODEL_IDS = ['text-embedding-3-small'] as const
 
-type SharedEmbeddingCache = {
-  get(model: string, text: string): Promise<number[] | null>
-  set(model: string, text: string, embedding: number[]): Promise<void>
-}
+/**
+ * Bound on the in-flight Promise dedup map. Concurrent embedQuery calls for
+ * the same text share a single API round-trip; this cap keeps the map from
+ * growing without bound under sustained load. FIFO eviction by insertion
+ * order — the oldest pending key is evicted before insertion when full. The
+ * pending Promise is not aborted; only its dedup slot is freed.
+ */
+const MAX_INFLIGHT = 1000
 
 type ProviderEmbeddingsOptions = {
   embedFn?: typeof embed
   embedManyFn?: typeof embedMany
-  getCache?: () => SharedEmbeddingCache | null
   modelId: string
   /**
    * `runPromise` is required: callers must yield `RuntimeBridgeTag` in their
@@ -67,7 +70,6 @@ function tryEmbeddingPromise<A>(
 export class ProviderEmbeddings {
   private readonly embedFn: typeof embed
   private readonly embedManyFn: typeof embedMany
-  private readonly getCache: () => SharedEmbeddingCache | null
   private readonly resolvedModelId: string
   private _model: ReturnType<typeof resolveEmbeddingModel> | null = null
   /** In-flight dedup: concurrent embedQuery calls for the same text share one API round-trip. */
@@ -78,7 +80,6 @@ export class ProviderEmbeddings {
   constructor(options: ProviderEmbeddingsOptions) {
     this.embedFn = options.embedFn ?? embed
     this.embedManyFn = options.embedManyFn ?? embedMany
-    this.getCache = options.getCache ?? (() => null)
     this.resolvedModelId = options.modelId
     this.runPromise = options.runPromise
   }
@@ -94,11 +95,12 @@ export class ProviderEmbeddings {
     return this._model
   }
 
-  private loadCachedEmbedding(text: string): Promise<number[] | null> {
-    const redisCache = this.getCache()
-    if (!redisCache) return Promise.resolve(null)
-
-    return redisCache.get(this.getModelId(), text)
+  private rememberInflight(key: string, promise: Promise<number[]>): void {
+    if (this.inflightEmbeddings.size >= MAX_INFLIGHT) {
+      const oldest = this.inflightEmbeddings.keys().next().value
+      if (oldest !== undefined) this.inflightEmbeddings.delete(oldest)
+    }
+    this.inflightEmbeddings.set(key, promise)
   }
 
   embedQuery(text: string): Promise<number[]> {
@@ -110,7 +112,11 @@ export class ProviderEmbeddings {
     if (pending) return pending
 
     const promise = this.runPromise(this.executeEmbedQueryEffect(input))
-    this.inflightEmbeddings.set(dedupKey, promise)
+    this.rememberInflight(dedupKey, promise)
+    // `.then(_, _)` (not `.finally`): the dual-handler form swallows the
+    // rejection on the cleanup branch so it never surfaces as an unhandled
+    // rejection. The original promise is still returned to the caller, who
+    // is responsible for reacting to its rejection.
     void promise.then(
       () => this.inflightEmbeddings.delete(dedupKey),
       () => this.inflightEmbeddings.delete(dedupKey),
@@ -122,24 +128,10 @@ export class ProviderEmbeddings {
   private executeEmbedQueryEffect(input: string): Effect.Effect<number[], EmbeddingProviderError> {
     return Effect.gen(
       function* (this: ProviderEmbeddings) {
-        const cached = yield* tryEmbeddingPromise('Failed to load cached query embedding.', () =>
-          this.loadCachedEmbedding(input),
-        )
-        if (cached) {
-          return cached
-        }
-
         const result = yield* tryEmbeddingPromise('Failed to generate query embedding.', () =>
           this.embedFn({ model: this.getModel(), value: input, maxRetries: 2 }),
         )
-        const embedding = normalizeEmbedding(result.embedding)
-
-        const redisCache = this.getCache()
-        if (redisCache) {
-          void redisCache.set(this.getModelId(), input, embedding)
-        }
-
-        return embedding
+        return normalizeEmbedding(result.embedding)
       }.bind(this),
     ).pipe(Effect.withSpan('ProviderEmbeddings.executeEmbedQuery'))
   }
@@ -166,45 +158,12 @@ export class ProviderEmbeddings {
   ): Effect.Effect<number[][], EmbeddingProviderError> {
     return Effect.gen(
       function* (this: ProviderEmbeddings) {
-        const embeddingsByText = new Map<string, number[]>()
-        let missingTexts = [...uniqueTexts]
-        const redisCache = this.getCache()
-        const redisResults =
-          redisCache && missingTexts.length > 0
-            ? yield* Effect.all(
-                missingTexts.map((text) =>
-                  tryEmbeddingPromise('Failed to load cached document embedding.', () =>
-                    redisCache.get(this.getModelId(), text),
-                  ).pipe(Effect.map((embedding) => ({ text, embedding }))),
-                ),
-              )
-            : ([] as Array<{ text: string; embedding: number[] | null }>)
-
-        if (redisCache && missingTexts.length > 0) {
-          missingTexts = []
-          for (const result of redisResults) {
-            if (!result.embedding) {
-              missingTexts.push(result.text)
-              continue
-            }
-
-            embeddingsByText.set(result.text, result.embedding)
-          }
-        }
-
-        if (missingTexts.length === 0) {
-          return normalized.map((text) => (text ? (embeddingsByText.get(text) ?? []) : []))
-        }
-
         const result = yield* tryEmbeddingPromise('Failed to generate document embeddings.', () =>
-          this.embedManyFn({ model: this.getModel(), values: missingTexts, maxRetries: 2 }),
+          this.embedManyFn({ model: this.getModel(), values: uniqueTexts, maxRetries: 2 }),
         )
-        missingTexts.forEach((text, index) => {
-          const embedding = normalizeEmbedding(result.embeddings[index] ?? [])
-          embeddingsByText.set(text, embedding)
-          if (redisCache) {
-            void redisCache.set(this.getModelId(), text, embedding)
-          }
+        const embeddingsByText = new Map<string, number[]>()
+        uniqueTexts.forEach((text, index) => {
+          embeddingsByText.set(text, normalizeEmbedding(result.embeddings[index] ?? []))
         })
 
         return normalized.map((text) => (text ? (embeddingsByText.get(text) ?? []) : []))

package/src/redis/index.ts

@@ -2,6 +2,7 @@ import { createRedisConnectionManager } from './connection'
 import type { RedisConnectionManager } from './connection'
 export { DEFAULT_REDIS_OPTIONS, type RedisConnectionLogger } from './connection'
 export { withOrgMemoryLock, withOrgMemoryLockEffect } from './org-memory-lock'
+export { RedisJsonCache, RedisJsonCacheError, RedisJsonCacheLive, RedisJsonCacheTag } from './redis-cache'
 export { withLeaseLock } from './redis-lease-lock'
 export {
   createThreadResumableContext,

package/src/redis/redis-cache.ts

@@ -0,0 +1,176 @@
+import { Context, Effect, Layer, Schema } from 'effect'
+import type IORedis from 'ioredis'
+
+import { RuntimeBridgeTag } from '../ai-gateway/ai-gateway'
+import { ERROR_TAGS } from '../effect/errors'
+import { RedisServiceTag } from '../effect/services'
+
+/**
+ * Bound on the in-flight Promise dedup map. Keeps memory usage predictable
+ * under sustained load — concurrent identical loads still share their
+ * Promise, but the map cannot grow without bound when ttls are long.
+ */
+const MAX_INFLIGHT = 1000
+
+export class RedisJsonCacheError extends Schema.TaggedErrorClass<RedisJsonCacheError>()(
+  ERROR_TAGS.RedisJsonCacheError,
+  {
+    message: Schema.String,
+    operation: Schema.Literals(['get', 'set', 'del', 'load']),
+    cause: Schema.optional(Schema.Defect),
+  },
+) {}
+
+/**
+ * Schema accepted by `RedisJsonCache` methods. The codec must be fully
+ * self-contained — no remaining decoding/encoding services — so that
+ * `Schema.fromJsonString(...)` resolves at runtime without additional
+ * context. `Schema.Codec<T, E, never, never>` is the canonical "ready to
+ * run" form.
+ */
+type RedisJsonCacheCodec<T> = Schema.Codec<T, unknown, never, never>
+
+/**
+ * Generic Redis-backed JSON cache keyed by string. The codec is supplied per
+ * call so a single cache instance can hold values of multiple shapes (the
+ * caller is responsible for namespacing keys correctly). Errors propagate as
+ * `RedisJsonCacheError`; this cache deliberately does not fail-soft.
+ */
+export class RedisJsonCache {
+  /** In-flight dedup: concurrent loads for the same key share a single round-trip. */
+  private readonly inflight = new Map<string, Promise<unknown>>()
+
+  constructor(
+    private readonly redis: IORedis,
+    private readonly runPromise: <A, E>(effect: Effect.Effect<A, E>) => Promise<A>,
+  ) {}
+
+  get<T>(key: string, schema: RedisJsonCacheCodec<T>): Promise<T | null> {
+    return this.runPromise(this.getEffect(key, schema))
+  }
+
+  set<T>(key: string, value: T, ttlSeconds: number, schema: RedisJsonCacheCodec<T>): Promise<void> {
+    return this.runPromise(this.setEffect(key, value, ttlSeconds, schema))
+  }
+
+  del(key: string): Promise<void> {
+    return this.runPromise(this.delEffect(key))
+  }
+
+  /**
+   * Read-through cache. On miss, invokes `loader` and stores the result with
+   * the supplied TTL. Concurrent calls with the same key share a single
+   * loader Promise via the in-flight map. The map is bounded — when full,
+   * the oldest entry (FIFO insertion order) is evicted before insert.
+   */
+  getOrSet<T>(key: string, ttlSeconds: number, schema: RedisJsonCacheCodec<T>, loader: () => Promise<T>): Promise<T> {
+    const existing = this.inflight.get(key) as Promise<T> | undefined
+    if (existing) return existing
+
+    const promise = this.runPromise(this.getOrSetEffect(key, ttlSeconds, schema, loader))
+    if (this.inflight.size >= MAX_INFLIGHT) {
+      const oldest = this.inflight.keys().next().value
+      if (oldest !== undefined) this.inflight.delete(oldest)
+    }
+    this.inflight.set(key, promise)
+    // `.then(_, _)` (not `.finally`): the dual-handler form swallows the
+    // rejection on the cleanup branch so it never surfaces as an unhandled
+    // rejection. The original promise is still returned to the caller, who
+    // is responsible for reacting to its rejection.
+    void promise.then(
+      () => this.inflight.delete(key),
+      () => this.inflight.delete(key),
+    )
+
+    return promise
+  }
+
+  private getOrSetEffect<T>(
+    key: string,
+    ttlSeconds: number,
+    schema: RedisJsonCacheCodec<T>,
+    loader: () => Promise<T>,
+  ): Effect.Effect<T, RedisJsonCacheError> {
+    return Effect.gen(
+      function* (this: RedisJsonCache) {
+        const cached = yield* this.getEffect(key, schema)
+        if (cached !== null) return cached
+
+        const fresh = yield* Effect.tryPromise({
+          try: () => loader(),
+          catch: (cause) =>
+            new RedisJsonCacheError({ message: 'Redis JSON cache loader failed.', operation: 'load', cause }),
+        })
+        yield* this.setEffect(key, fresh, ttlSeconds, schema)
+        return fresh
+      }.bind(this),
+    )
+  }
+
+  private getEffect<T>(key: string, schema: RedisJsonCacheCodec<T>): Effect.Effect<T | null, RedisJsonCacheError> {
+    const redis = this.redis
+    const decode = Schema.decodeUnknownSync(Schema.fromJsonString(schema)) as (raw: string) => T
+    return Effect.gen(function* () {
+      const cached = yield* Effect.tryPromise({
+        try: () => redis.getBuffer(key),
+        catch: (cause) => new RedisJsonCacheError({ message: 'Redis JSON cache get failed.', operation: 'get', cause }),
+      })
+
+      if (!cached) return null
+
+      return yield* Effect.try({
+        try: () => decode(cached.toString()),
+        catch: (cause) =>
+          new RedisJsonCacheError({ message: 'Redis JSON cache parse failed.', operation: 'get', cause }),
+      })
+    })
+  }
+
+  private setEffect<T>(
+    key: string,
+    value: T,
+    ttlSeconds: number,
+    schema: RedisJsonCacheCodec<T>,
+  ): Effect.Effect<void, RedisJsonCacheError> {
+    const redis = this.redis
+    const encode = Schema.encodeSync(Schema.fromJsonString(schema)) as (input: T) => string
+    return Effect.gen(function* () {
+      const serialized = yield* Effect.try({
+        try: () => encode(value),
+        catch: (cause) =>
+          new RedisJsonCacheError({ message: 'Redis JSON cache serialization failed.', operation: 'set', cause }),
+      })
+      yield* Effect.tryPromise({
+        try: () => redis.set(key, serialized, 'EX', ttlSeconds),
+        catch: (cause) => new RedisJsonCacheError({ message: 'Redis JSON cache set failed.', operation: 'set', cause }),
+      })
+    }).pipe(Effect.asVoid)
+  }
+
+  private delEffect(key: string): Effect.Effect<void, RedisJsonCacheError> {
+    const redis = this.redis
+    return Effect.tryPromise({
+      try: () => redis.del(key),
+      catch: (cause) => new RedisJsonCacheError({ message: 'Redis JSON cache del failed.', operation: 'del', cause }),
+    }).pipe(Effect.asVoid)
+  }
+}
+
+export class RedisJsonCacheTag extends Context.Service<RedisJsonCacheTag, RedisJsonCache>()(
+  '@lota-sdk/core/RedisJsonCache',
+) {}
+
+/**
+ * `RedisJsonCache` needs a `runPromise` so its Promise-returning surface stays
+ * grounded in the host runtime. We yield `RuntimeBridgeTag` once at layer
+ * construction and capture the bridged `runPromise`; consumers never
+ * prop-drill it.
+ */
+export const RedisJsonCacheLive = Layer.effect(
+  RedisJsonCacheTag,
+  Effect.gen(function* () {
+    const redis = yield* RedisServiceTag
+    const bridge = yield* RuntimeBridgeTag
+    return new RedisJsonCache(redis.getConnection(), bridge.runPromise)
+  }),
+)

package/src/runtime/domain-layer.ts

@@ -11,9 +11,9 @@ import type { Layer as LayerType } from 'effect'
 import { Layer } from 'effect'
 
 import type { AiGatewayModelsTag, AiGatewayTag, RuntimeBridgeTag } from '../ai-gateway/ai-gateway'
-import { EmbeddingCacheLive } from '../ai/embedding-cache'
 import type { buildInfrastructureLayer } from '../effect/layers'
 import { LotaQueuesLive } from '../queues/queues.service'
+import { RedisJsonCacheLive } from '../redis/redis-cache'
 import { SharedThreadStreamSubscriberLive } from '../redis/stream-context'
 import { AgentActivityServiceLive } from '../services/agent-activity.service'
 import { AgentExecutorServiceLive } from '../services/agent-executor.service'
@@ -113,7 +113,7 @@ export function buildDomainServiceLayer(infrastructureLayer: InfrastructureLayer
   )
   const ctx0 = Layer.mergeAll(
     baseCtx,
-    provide(Layer.mergeAll(EmbeddingCacheLive, FirecrawlLive, HelperModelLive), baseCtx),
+    provide(Layer.mergeAll(FirecrawlLive, HelperModelLive, RedisJsonCacheLive), baseCtx),
     tier0,
   )
 

package/src/runtime/runtime-config.ts

@@ -237,16 +237,10 @@ export const LotaRuntimeConfigSchema = z.object({
   memory: z
     .object({
       searchK: z.coerce.number().int().positive().default(6),
-      embeddingCacheTtlSeconds: z.coerce.number().int().positive().default(7200),
       rerankerStrategy: MemoryRerankerStrategySchema.default('rerank'),
       rerankerModelId: z.string().trim().min(1).default(AI_GATEWAY_FAST_RERANK_MODEL_ID),
     })
-    .default({
-      searchK: 6,
-      embeddingCacheTtlSeconds: 7200,
-      rerankerStrategy: 'rerank',
-      rerankerModelId: AI_GATEWAY_FAST_RERANK_MODEL_ID,
-    }),
+    .default({ searchK: 6, rerankerStrategy: 'rerank', rerankerModelId: AI_GATEWAY_FAST_RERANK_MODEL_ID }),
   threads: threadConfigSchema.default({}),
   agents: agentsConfigSchema,
   toolProviders: z.custom<ToolSet>(isToolSet, { error: 'toolProviders must be a tool registry object' }).optional(),

package/src/services/plan/plan-deadline.service.ts

@@ -393,7 +393,7 @@ export function makePlanDeadlineService(deps: PlanDeadlineDeps) {
       return
     }
 
-    const runCache = new Map<string, PlanRunRecord>()
+    const runLookup = new Map<string, PlanRunRecord>()
 
     const handleEntry = (entry: (typeof sweep.entries)[number]): Effect.Effect<void, PlanDeadlineError> =>
       Effect.gen(function* () {
@@ -403,9 +403,9 @@ export function makePlanDeadlineService(deps: PlanDeadlineDeps) {
         }
 
         const runIdStr = recordIdToString(entry.nodeRun.runId, TABLES.PLAN_RUN)
-        const cachedRun = runCache.get(runIdStr)
+        const existing = runLookup.get(runIdStr)
         const run =
-          cachedRun ??
+          existing ??
           (yield* db
             .findOne(TABLES.PLAN_RUN, { id: ensureRecordId(entry.nodeRun.runId, TABLES.PLAN_RUN) }, PlanRunSchema)
             .pipe(
@@ -414,7 +414,7 @@ export function makePlanDeadlineService(deps: PlanDeadlineDeps) {
         if (!run) {
           return
         }
-        runCache.set(runIdStr, run)
+        runLookup.set(runIdStr, run)
 
         const dedupeKeyBase = `plan-deadline:${runIdStr}:${entry.nodeRun.nodeId}`
         const actionEffect =

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import { OPENROUTER_STRUCTURED_HELPER_MODEL_ID } from '../config/model-constants'
 import { resolveHelperAgentOptions } from './helper-agent-options'
 
@@ -33,7 +32,6 @@ export function makeContextCompactionAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'context-compaction',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       ...resolveHelperAgentOptions(options, { instructions: CONTEXT_COMPACTION_PROMPT }),
     })
 }

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
@@ -34,7 +33,6 @@ export function makeMemoryRerankerAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'memory-reranker',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options),
     })

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -54,7 +53,6 @@ export function makeOrgMemoryAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'org-memory',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options),
     })

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import type { ResolvedAgentConfig } from '../config/agent-defaults'
 import { OPENROUTER_STRUCTURED_HELPER_MODEL_ID } from '../config/model-constants'
 import { resolveHelperAgentOptions } from './helper-agent-options'
@@ -78,7 +77,6 @@ export function makeRecentActivityTitleRefinerAgentFactory(models: AiGatewayMode
     new ToolLoopAgent({
       id: 'recent-activity-title-refiner',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       ...resolveHelperAgentOptions(options, {
         instructions: buildRecentActivityTitleRefinerPrompt(agentConfig),
         maxOutputTokens: RECENT_ACTIVITY_TITLE_MAX_TOKENS,

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -29,7 +28,6 @@ export function makeRegularChatMemoryDigestAgentFactory(models: AiGatewayModels)
     new ToolLoopAgent({
       id: 'regular-chat-memory-digest',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options, {
         instructions: regularChatMemoryDigestPrompt,

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

@@ -3,7 +3,6 @@ import { ToolLoopAgent } from 'ai'
 import { z } from 'zod'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -47,7 +46,6 @@ export function makeSkillExtractorAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'skill-extractor',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options, {
         instructions: skillExtractorPrompt,

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

@@ -3,7 +3,6 @@ import { ToolLoopAgent } from 'ai'
 import { z } from 'zod'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_STRUCTURED_HELPER_MODEL_ID,
   OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
@@ -71,7 +70,6 @@ export function makeSkillManagerAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'skill-manager',
       model: models.chatModel(OPENROUTER_STRUCTURED_HELPER_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_HIGH_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options, {
         instructions: skillManagerPrompt,

package/src/system-agents/thread-router.agent.ts

@@ -3,7 +3,6 @@ import { Effect } from 'effect'
 import { z } from 'zod'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import type { ResolvedAgentConfig } from '../config/agent-defaults'
 import { chatLogger } from '../config/logger'
 import type { ValidationError } from '../effect/errors'
@@ -182,7 +181,6 @@ function generateRouterObjectEffect<TSchema extends z.ZodTypeAny>(params: {
     try: () =>
       generateObject({
         model: params.aiGatewayModels.chatModel(modelId),
-        headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
         schema: params.schema,
         system: params.system,
         prompt: params.prompt,

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

@@ -2,7 +2,6 @@ import type { CreateHelperToolLoopAgentOptions } from '@lota-sdk/shared'
 import { ToolLoopAgent } from 'ai'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayDirectCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_FAST_REASONING_MODEL_ID,
   OPENROUTER_MINIMAL_REASONING_PROVIDER_OPTIONS,
@@ -35,7 +34,6 @@ export function makeThreadTitleGeneratorAgentFactory(models: AiGatewayModels) {
     new ToolLoopAgent({
       id: 'thread-title-generator',
       model: models.chatModel(OPENROUTER_FAST_REASONING_MODEL_ID),
-      headers: buildAiGatewayDirectCacheHeaders('lota-sdk'),
       providerOptions: OPENROUTER_MINIMAL_REASONING_PROVIDER_OPTIONS,
       ...resolveHelperAgentOptions(options, {
         instructions: THREAD_TITLE_GENERATOR_PROMPT,

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

@@ -3,7 +3,6 @@ import type { Effect } from 'effect'
 
 import type { AiGatewayModels } from '../ai-gateway/ai-gateway'
 import { aiGatewayChatModel, getDefaultAiGatewayRunPromise } from '../ai-gateway/ai-gateway'
-import { buildAiGatewayStrictSemanticCacheHeaders } from '../ai-gateway/cache-headers'
 import {
   OPENROUTER_FAST_REASONING_MODEL_ID,
   OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
@@ -28,7 +27,6 @@ export const researchTopicTool = createDelegatedAgentToolWithContext<ResearchTop
       ? aiGatewayModels.chatModel(OPENROUTER_FAST_REASONING_MODEL_ID)
       : aiGatewayChatModel(OPENROUTER_FAST_REASONING_MODEL_ID),
   providerOptions: OPENROUTER_LOW_REASONING_PROVIDER_OPTIONS,
-  headers: buildAiGatewayStrictSemanticCacheHeaders('researchTopic'),
   instructions: RESEARCHER_PROMPT,
   createTools: ({ firecrawl, runPromise }) => ({
     searchWeb: searchWebTool.create({ firecrawl, runPromise }),

package/src/ai/embedding-cache.ts

@@ -1,127 +0,0 @@
-import { Context, Effect, Layer, Schema } from 'effect'
-import type IORedis from 'ioredis'
-
-import { RuntimeBridgeTag } from '../ai-gateway/ai-gateway'
-import { aiLogger } from '../config/logger'
-import { ERROR_TAGS } from '../effect/errors'
-import { RedisServiceTag, RuntimeConfigServiceTag } from '../effect/services'
-import { sha256Hex } from '../utils/crypto'
-
-export const DEFAULT_EMBEDDING_CACHE_TTL_SECONDS = 7200
-const EMBEDDING_CACHE_KEY_PREFIX = 'emb'
-const EmbeddingCacheJsonSchema = Schema.fromJsonString(Schema.Array(Schema.Number))
-
-class EmbeddingCacheError extends Schema.TaggedErrorClass<EmbeddingCacheError>()(ERROR_TAGS.EmbeddingCacheError, {
-  message: Schema.String,
-  operation: Schema.Literals(['get', 'set']),
-  cause: Schema.optional(Schema.Defect),
-}) {}
-
-function decodeEmbeddingCacheValue(raw: string): number[] {
-  return [...Schema.decodeUnknownSync(EmbeddingCacheJsonSchema)(raw)]
-}
-
-function encodeEmbeddingCacheValue(embedding: number[]): string {
-  return Schema.encodeSync(EmbeddingCacheJsonSchema)(embedding)
-}
-
-export class EmbeddingCache {
-  /** In-flight dedup: concurrent gets for the same key share a single Redis+API round-trip. */
-  private readonly inflight = new Map<string, Promise<number[] | null>>()
-
-  constructor(
-    private redis: IORedis,
-    private ttlSeconds: number,
-    private readonly runPromise: <A, E = never>(effect: Effect.Effect<A, E>) => Promise<A>,
-  ) {}
-
-  private buildKey(model: string, text: string): string {
-    const hash = sha256Hex(text)
-    return `${EMBEDDING_CACHE_KEY_PREFIX}:${model}:${hash}`
-  }
-
-  private fetchFromRedisEffect(key: string): Effect.Effect<number[] | null> {
-    const redis = this.redis
-
-    return Effect.gen(function* () {
-      const cached = yield* Effect.tryPromise({
-        try: () => redis.getBuffer(key),
-        catch: (cause) => new EmbeddingCacheError({ message: 'Embedding cache get failed.', operation: 'get', cause }),
-      })
-
-      if (!cached) return null
-
-      return yield* Effect.try({
-        try: () => decodeEmbeddingCacheValue(cached.toString()),
-        catch: (cause) =>
-          new EmbeddingCacheError({ message: 'Embedding cache parse failed.', operation: 'get', cause }),
-      })
-    }).pipe(
-      Effect.catchTag(ERROR_TAGS.EmbeddingCacheError, (error) => {
-        aiLogger.warn`${error.message}: ${error.cause}`
-        return Effect.succeed<number[] | null>(null)
-      }),
-    )
-  }
-
-  get(model: string, text: string): Promise<number[] | null> {
-    const key = this.buildKey(model, text)
-
-    const pending = this.inflight.get(key)
-    if (pending) return pending
-
-    const promise = this.runPromise(this.fetchFromRedisEffect(key))
-    this.inflight.set(key, promise)
-    void promise.finally(() => this.inflight.delete(key))
-
-    return promise
-  }
-
-  private setEffect(model: string, text: string, embedding: number[]): Effect.Effect<void> {
-    const redis = this.redis
-    const ttlSeconds = this.ttlSeconds
-    const key = this.buildKey(model, text)
-
-    return Effect.gen(function* () {
-      const serialized = yield* Effect.try({
-        try: () => encodeEmbeddingCacheValue(embedding),
-        catch: (cause) =>
-          new EmbeddingCacheError({ message: 'Embedding cache serialization failed.', operation: 'set', cause }),
-      })
-      yield* Effect.tryPromise({
-        try: () => redis.set(key, serialized, 'EX', ttlSeconds),
-        catch: (cause) => new EmbeddingCacheError({ message: 'Embedding cache set failed.', operation: 'set', cause }),
-      })
-    }).pipe(
-      Effect.asVoid,
-      Effect.catchTag(ERROR_TAGS.EmbeddingCacheError, (error) => {
-        aiLogger.warn`${error.message}: ${error.cause}`
-        return Effect.void
-      }),
-    )
-  }
-
-  set(model: string, text: string, embedding: number[]): Promise<void> {
-    return this.runPromise(this.setEffect(model, text, embedding))
-  }
-}
-
-export class EmbeddingCacheTag extends Context.Service<EmbeddingCacheTag, EmbeddingCache>()(
-  '@lota-sdk/core/EmbeddingCache',
-) {}
-
-/**
- * `EmbeddingCache` needs a `runPromise` to convert internal Effect chains into
- * the Promise API that AI SDK `embed` / `embedMany` expect. The layer yields
- * `RuntimeBridgeTag` once and captures it — callers never prop-drill
- * `runPromise` through this module.
- */
-export const EmbeddingCacheLive = Layer.effect(
-  EmbeddingCacheTag,
-  Effect.gen(function* () {
-    const redis = yield* RedisServiceTag
-    const config = yield* RuntimeConfigServiceTag
-    const bridge = yield* RuntimeBridgeTag
-    return new EmbeddingCache(redis.getConnection(), config.memory.embeddingCacheTtlSeconds, bridge.runPromise)
-  }),
-)

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

@@ -1,42 +0,0 @@
-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 const AI_GATEWAY_STRICT_SEMANTIC_CACHE_THRESHOLD = 0.975
-
-export type AiGatewayCacheType = 'direct' | 'semantic'
-
-export function toAiGatewayCacheKeyPart(value: string): string {
-  const normalized = value
-    .trim()
-    .toLowerCase()
-    .replace(/[^a-z0-9:_-]+/g, '-')
-    .replace(/-+/g, '-')
-  return normalized.replace(/^-+|-+$/g, '') || 'request'
-}
-
-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')
-}