@ax-llm/ax 19.0.27 → 19.0.29

package/index.d.cts

@@ -1004,6 +1004,18 @@ interface AxAIServiceImpl<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TCha
      * Returns cache operations to execute and the modified request.
      */
     prepareCachedChatReq?(req: Readonly<AxInternalChatRequest<TModel>>, options: Readonly<AxAIServiceOptions>, existingCacheName?: string): Promise<AxPreparedChatRequest<TChatRequest>>;
+    /**
+     * Optional: Return tool/function-call state that becomes part of the
+     * provider's immutable explicit-cache prefix.
+     *
+     * This is used for cache identity and token estimation when a provider
+     * treats tool state as cached prefix state even if the generic request does
+     * not mark functions with `cache: true`.
+     */
+    getContextCacheToolState?(req: Readonly<AxInternalChatRequest<TModel>>, options: Readonly<AxAIServiceOptions>): {
+        functions?: AxChatRequest['functions'];
+        functionCall?: AxChatRequest['functionCall'];
+    } | undefined;
     /**
      * Optional: Build a context cache creation operation.
      * Called when a new cache needs to be created from the request.
@@ -2942,6 +2954,7 @@ declare class AxGen<IN = any, OUT extends AxGenOut = any> extends AxProgram<IN,
     setInstruction(instruction: string): void;
     getInstruction(): string | undefined;
     clearInstruction(): void;
+    private getEffectiveContextCache;
     private renderPromptWithMetricsForInternalUse;
     private renderPromptForInternalUse;
     /** @internal */
@@ -3172,6 +3185,8 @@ declare class AxPromptTemplate {
      */
     private buildOutputFieldsSection;
     private renderSingleValueUserContent;
+    private formatUserContent;
+    private buildLegacyMultimodalExampleMessage;
     private renderInternal;
     render: <T = any>(values: T | ReadonlyArray<AxMessage<T>>, options: Readonly<{
         skipSystemPrompt?: boolean;
@@ -3753,6 +3768,7 @@ declare class AxBaseAI<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TChatRe
     private corsProxy?;
     private retry?;
     private customLabels?;
+    private contextCache?;
     private modelInfo;
     private modelUsage?;
     private embedModelUsage?;
@@ -3833,6 +3849,7 @@ declare class AxBaseAI<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TChatRe
      * Use an existing cache by name to prepare the chat request.
      */
     private useCacheByName;
+    private getContextCacheToolState;
     /**
      * Execute a context cache operation (create/update/delete).
      */
@@ -10271,8 +10288,12 @@ type AxAgentOptions<IN extends AxGenIn = AxGenIn> = Omit<AxProgramForwardOptions
     maxBatchedLlmQueryConcurrency?: number;
     /** Maximum Actor turns before forcing Responder (default: 10). */
     maxTurns?: number;
+    /** Maximum characters to keep from runtime output and console/log replay. */
+    maxRuntimeChars?: number;
     /** Context replay, checkpointing, and runtime-state policy. */
     contextPolicy?: AxContextPolicyConfig;
+    /** Default options for the internal checkpoint summarizer. */
+    summarizerOptions?: Omit<AxProgramForwardOptions<string>, 'functions'>;
     /** Output field names the Actor should produce (in addition to javascriptCode). */
     actorFields?: string[];
     /**
@@ -10684,7 +10705,7 @@ interface AxCodeSession {
  *   replaying their full code blocks. Reliability-first
  *   defaults still preserve recent evidence before deleting older low-value steps.
  *   Best when token pressure matters more than raw replay detail.
- * - `checkpointed`: Keep full replay until the rendered actor prompt crosses a threshold, then
+ * - `checkpointed`: Keep full replay until the rendered actor prompt grows beyond the selected budget, then
  *   replace older successful history with a checkpoint summary while keeping recent
  *   actions and unresolved errors fully visible. Best when you want conservative,
  *   debugging-friendly replay until prompt pressure becomes real.
@@ -10702,7 +10723,7 @@ interface AxContextPolicyConfig {
      * - `full`: prefer raw replay of earlier actions
      * - `adaptive`: balance replay detail with checkpoint compression while keeping more recent evidence visible
      * - `lean`: prefer live state + compact summaries over raw replay detail
-     * - `checkpointed`: keep full replay until the rendered actor prompt crosses a threshold, then replace older successful turns with a checkpoint summary
+     * - `checkpointed`: keep full replay until the rendered actor prompt grows beyond the selected budget, then replace older successful turns with a checkpoint summary
      */
     preset?: AxContextPolicyPreset;
     /** Overall prompt budget and compression aggressiveness. */
@@ -10726,8 +10747,12 @@ interface AxRLMConfig {
     maxBatchedLlmQueryConcurrency?: number;
     /** Maximum Actor turns before forcing Responder (default: 10). */
     maxTurns?: number;
+    /** Maximum characters to keep from runtime output and console/log replay (default: 3000). */
+    maxRuntimeChars?: number;
     /** Context replay, checkpointing, and runtime-state policy. */
     contextPolicy?: AxContextPolicyConfig;
+    /** Default options for the internal checkpoint summarizer. */
+    summarizerOptions?: Omit<AxProgramForwardOptions<string>, 'functions'>;
     /** Output field names the Actor should produce (in addition to javascriptCode). */
     actorFields?: string[];
     /**

package/index.d.ts

@@ -1004,6 +1004,18 @@ interface AxAIServiceImpl<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TCha
      * Returns cache operations to execute and the modified request.
      */
     prepareCachedChatReq?(req: Readonly<AxInternalChatRequest<TModel>>, options: Readonly<AxAIServiceOptions>, existingCacheName?: string): Promise<AxPreparedChatRequest<TChatRequest>>;
+    /**
+     * Optional: Return tool/function-call state that becomes part of the
+     * provider's immutable explicit-cache prefix.
+     *
+     * This is used for cache identity and token estimation when a provider
+     * treats tool state as cached prefix state even if the generic request does
+     * not mark functions with `cache: true`.
+     */
+    getContextCacheToolState?(req: Readonly<AxInternalChatRequest<TModel>>, options: Readonly<AxAIServiceOptions>): {
+        functions?: AxChatRequest['functions'];
+        functionCall?: AxChatRequest['functionCall'];
+    } | undefined;
     /**
      * Optional: Build a context cache creation operation.
      * Called when a new cache needs to be created from the request.
@@ -2942,6 +2954,7 @@ declare class AxGen<IN = any, OUT extends AxGenOut = any> extends AxProgram<IN,
     setInstruction(instruction: string): void;
     getInstruction(): string | undefined;
     clearInstruction(): void;
+    private getEffectiveContextCache;
     private renderPromptWithMetricsForInternalUse;
     private renderPromptForInternalUse;
     /** @internal */
@@ -3172,6 +3185,8 @@ declare class AxPromptTemplate {
      */
     private buildOutputFieldsSection;
     private renderSingleValueUserContent;
+    private formatUserContent;
+    private buildLegacyMultimodalExampleMessage;
     private renderInternal;
     render: <T = any>(values: T | ReadonlyArray<AxMessage<T>>, options: Readonly<{
         skipSystemPrompt?: boolean;
@@ -3753,6 +3768,7 @@ declare class AxBaseAI<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TChatRe
     private corsProxy?;
     private retry?;
     private customLabels?;
+    private contextCache?;
     private modelInfo;
     private modelUsage?;
     private embedModelUsage?;
@@ -3833,6 +3849,7 @@ declare class AxBaseAI<TModel, TEmbedModel, TChatRequest, TEmbedRequest, TChatRe
      * Use an existing cache by name to prepare the chat request.
      */
     private useCacheByName;
+    private getContextCacheToolState;
     /**
      * Execute a context cache operation (create/update/delete).
      */
@@ -10271,8 +10288,12 @@ type AxAgentOptions<IN extends AxGenIn = AxGenIn> = Omit<AxProgramForwardOptions
     maxBatchedLlmQueryConcurrency?: number;
     /** Maximum Actor turns before forcing Responder (default: 10). */
     maxTurns?: number;
+    /** Maximum characters to keep from runtime output and console/log replay. */
+    maxRuntimeChars?: number;
     /** Context replay, checkpointing, and runtime-state policy. */
     contextPolicy?: AxContextPolicyConfig;
+    /** Default options for the internal checkpoint summarizer. */
+    summarizerOptions?: Omit<AxProgramForwardOptions<string>, 'functions'>;
     /** Output field names the Actor should produce (in addition to javascriptCode). */
     actorFields?: string[];
     /**
@@ -10684,7 +10705,7 @@ interface AxCodeSession {
  *   replaying their full code blocks. Reliability-first
  *   defaults still preserve recent evidence before deleting older low-value steps.
  *   Best when token pressure matters more than raw replay detail.
- * - `checkpointed`: Keep full replay until the rendered actor prompt crosses a threshold, then
+ * - `checkpointed`: Keep full replay until the rendered actor prompt grows beyond the selected budget, then
  *   replace older successful history with a checkpoint summary while keeping recent
  *   actions and unresolved errors fully visible. Best when you want conservative,
  *   debugging-friendly replay until prompt pressure becomes real.
@@ -10702,7 +10723,7 @@ interface AxContextPolicyConfig {
      * - `full`: prefer raw replay of earlier actions
      * - `adaptive`: balance replay detail with checkpoint compression while keeping more recent evidence visible
      * - `lean`: prefer live state + compact summaries over raw replay detail
-     * - `checkpointed`: keep full replay until the rendered actor prompt crosses a threshold, then replace older successful turns with a checkpoint summary
+     * - `checkpointed`: keep full replay until the rendered actor prompt grows beyond the selected budget, then replace older successful turns with a checkpoint summary
      */
     preset?: AxContextPolicyPreset;
     /** Overall prompt budget and compression aggressiveness. */
@@ -10726,8 +10747,12 @@ interface AxRLMConfig {
     maxBatchedLlmQueryConcurrency?: number;
     /** Maximum Actor turns before forcing Responder (default: 10). */
     maxTurns?: number;
+    /** Maximum characters to keep from runtime output and console/log replay (default: 3000). */
+    maxRuntimeChars?: number;
     /** Context replay, checkpointing, and runtime-state policy. */
     contextPolicy?: AxContextPolicyConfig;
+    /** Default options for the internal checkpoint summarizer. */
+    summarizerOptions?: Omit<AxProgramForwardOptions<string>, 'functions'>;
     /** Output field names the Actor should produce (in addition to javascriptCode). */
     actorFields?: string[];
     /**