@kuralle-agents/core 0.7.1 → 0.7.2

package/dist/index.d.ts

@@ -55,6 +55,8 @@ export { resolveAgentWorkspace, type ResolvedAgentWorkspace, } from './runtime/r
 export { DEFAULT_CONTEXT_BUDGET, VOICE_CONTEXT_BUDGET, computeMessageHistoryBudget, truncateToTokenBudget, formatMemoryWithBudget, estimateTokenCount, ContextBudget, } from './runtime/ContextBudget.js';
 export type { ContextBudgetConfig } from './runtime/ContextBudget.js';
 export { TokenAccumulator } from './runtime/TokenAccumulator.js';
+export { applyPromptCache, applyAnthropicCacheControl, buildOpenAIResponsesProviderOptions, isAnthropicLanguageModel, isOpenAIResponsesModel, } from './runtime/promptCache.js';
+export type { AnthropicCacheTtl, OpenAIResponsesCompactOptions } from './runtime/promptCache.js';
 export { handoffFilters, removeToolHistory, keepRecentMessages, removeKeys, composeFilters, } from './runtime/handoffFilters.js';
 export type { HandoffInputFilter, HandoffInputData, HandoffInputResult, } from './runtime/handoffFilters.js';
 export { DefaultToolExecutor, DefaultConversationState, DefaultConversationEventLog, DefaultAgentStateController, createFoundation, } from './foundation/index.js';

package/dist/index.js

@@ -34,6 +34,7 @@ export { wireWorkingMemory, loadWorkingMemoryBlocks, formatWorkingMemorySection,
 export { resolveAgentWorkspace, } from './runtime/resolveAgentWorkspace.js';
 export { DEFAULT_CONTEXT_BUDGET, VOICE_CONTEXT_BUDGET, computeMessageHistoryBudget, truncateToTokenBudget, formatMemoryWithBudget, estimateTokenCount, ContextBudget, } from './runtime/ContextBudget.js';
 export { TokenAccumulator } from './runtime/TokenAccumulator.js';
+export { applyPromptCache, applyAnthropicCacheControl, buildOpenAIResponsesProviderOptions, isAnthropicLanguageModel, isOpenAIResponsesModel, } from './runtime/promptCache.js';
 export { handoffFilters, removeToolHistory, keepRecentMessages, removeKeys, composeFilters, } from './runtime/handoffFilters.js';
 export { DefaultToolExecutor, DefaultConversationState, DefaultConversationEventLog, DefaultAgentStateController, createFoundation, } from './foundation/index.js';
 export { CapabilityHost, ExtractionCapability, GuardrailCapability, AutoRetrieveCapability, PassThroughRefinement, PassThroughValidation, toGeminiDeclarations, toAISDKTools, } from './capabilities/index.js';

package/dist/runtime/channels/TextDriver.js

@@ -9,6 +9,7 @@ import { resolveMaxSteps } from '../policies/limits.js';
 import { speakWithHostControl } from './streaming/hostControlSpeak.js';
 import { resolveStreamMode } from './streaming/mode.js';
 import { appendGatherBlocks, resolveNodeGatherScope, runGatherPhase } from '../grounding/index.js';
+import { applyPromptCache } from '../promptCache.js';
 import { isFlowTransitionControlTool } from '../../flow/flowControlTools.js';
 import { resolveStructuredDecide } from '../../flow/choiceMatch.js';
 export class TextDriver {
@@ -50,12 +51,14 @@ export class TextDriver {
         const source = {
             async *[Symbol.asyncIterator]() {
                 for (let step = 0; step < maxSteps; step += 1) {
+                    const cached = applyPromptCache(model, ctx.session.id, messages);
                     const result = streamText({
                         model,
                         system,
-                        messages,
+                        messages: cached.messages,
                         tools: aiTools,
                         abortSignal: ctx.abortSignal,
+                        ...(cached.providerOptions ? { providerOptions: cached.providerOptions } : {}),
                     });
                     for await (const part of result.fullStream) {
                         if (part.type === 'text-delta') {

package/dist/runtime/channels/extractionTurn.js

@@ -2,6 +2,7 @@ import { streamText } from 'ai';
 import { executeModelToolCall, toolResultMessage } from './executeModelTool.js';
 import { buildToolSet } from '../../tools/effect/index.js';
 import { buildNodePrompt, composeSystem } from '../../flow/nodeBuilders.js';
+import { applyPromptCache } from '../promptCache.js';
 /**
  * Shared, NON-SPEAKING field extraction for `collect` nodes, used by every
  * ChannelDriver so text and voice behave identically. It runs the model with the
@@ -20,13 +21,15 @@ export async function runSilentExtraction(node, ctx, model, maxSteps) {
     const aiTools = resolveExtractionTools(node);
     const out = { text: '', toolResults: [] };
     for (let step = 0; step < maxSteps; step += 1) {
+        const cached = applyPromptCache(model, ctx.session.id, messages);
         const result = streamText({
             model,
             system,
-            messages,
+            messages: cached.messages,
             tools: aiTools,
             temperature: 0,
             abortSignal: ctx.abortSignal,
+            ...(cached.providerOptions ? { providerOptions: cached.providerOptions } : {}),
         });
         for await (const part of result.fullStream) {
             // Intentionally NOT handling 'text-delta' — extraction never speaks.

package/dist/runtime/promptCache.d.ts

@@ -23,7 +23,7 @@
  * but extends caching across short user idle periods + cross-session
  * within an hour.
  */
-import type { ModelMessage } from 'ai';
+import type { JSONValue, ModelMessage } from 'ai';
 export type AnthropicCacheTtl = '5m' | '1h';
 /**
  * Returns a SHALLOW-COPIED message list with cache_control markers
@@ -90,3 +90,19 @@ export interface OpenAIResponsesCompactOptions {
  * branch but ignored otherwise.
  */
 export declare function buildOpenAIResponsesProviderOptions(opts: OpenAIResponsesCompactOptions, sessionId: string): Record<string, unknown> | null;
+/**
+ * Single wiring point for provider prompt caching, called by the driver
+ * `streamText` sites. Gated by the conservative detectors so non-matching
+ * providers are untouched:
+ *   - Anthropic → applies `cache_control` breakpoints (caches the system+tools
+ *     prefix + recent history; ~75% input-cost + TTFT off multi-turn turns).
+ *   - OpenAI Responses → sets `promptCacheKey` (= sessionId, pins same-session
+ *     turns to one cache slot) + `truncation: 'auto'` overflow safety net.
+ * Returns the (possibly transformed) messages + an optional `providerOptions`
+ * bag ready to spread into `streamText`. Default-on: both are no-ops on
+ * providers that ignore the fields.
+ */
+export declare function applyPromptCache(model: unknown, sessionId: string, messages: ModelMessage[]): {
+    messages: ModelMessage[];
+    providerOptions?: Record<string, Record<string, JSONValue>>;
+};

package/dist/runtime/promptCache.js

@@ -110,6 +110,30 @@ export function buildOpenAIResponsesProviderOptions(opts, sessionId) {
     }
     return Object.keys(out).length > 0 ? out : null;
 }
+/**
+ * Single wiring point for provider prompt caching, called by the driver
+ * `streamText` sites. Gated by the conservative detectors so non-matching
+ * providers are untouched:
+ *   - Anthropic → applies `cache_control` breakpoints (caches the system+tools
+ *     prefix + recent history; ~75% input-cost + TTFT off multi-turn turns).
+ *   - OpenAI Responses → sets `promptCacheKey` (= sessionId, pins same-session
+ *     turns to one cache slot) + `truncation: 'auto'` overflow safety net.
+ * Returns the (possibly transformed) messages + an optional `providerOptions`
+ * bag ready to spread into `streamText`. Default-on: both are no-ops on
+ * providers that ignore the fields.
+ */
+export function applyPromptCache(model, sessionId, messages) {
+    const outMessages = isAnthropicLanguageModel(model)
+        ? applyAnthropicCacheControl(messages)
+        : messages;
+    if (isOpenAIResponsesModel(model)) {
+        const openai = buildOpenAIResponsesProviderOptions({ useSessionAsPromptCacheKey: true, truncationFallback: 'auto' }, sessionId);
+        if (openai) {
+            return { messages: outMessages, providerOptions: { openai: openai } };
+        }
+    }
+    return { messages: outMessages };
+}
 function withProviderOptions(msg, cacheControl) {
     const existing = msg.providerOptions ?? {};
     const existingAnthropic = existing.anthropic ?? {};

package/package.json

@@ -6,7 +6,7 @@
     "url": "git+https://github.com/kuralle/kuralle-agents.git",
     "directory": "packages/kuralle-core"
   },
-  "version": "0.7.1",
+  "version": "0.7.2",
   "description": "A framework for structured conversational AI agents",
   "publishConfig": {
     "access": "public"
@@ -103,7 +103,7 @@
     "typescript": "^5.3.0",
     "vitest": "^3.2.4",
     "zod": "^4.0.0",
-    "@kuralle-agents/realtime-audio": "0.7.1"
+    "@kuralle-agents/realtime-audio": "0.7.2"
   },
   "dependencies": {
     "chrono-node": "^2.6.0"