@jambonz/llm 0.6.9 → 0.6.11

package/dist/{types-C_yrWq_0.d.cts → types-C3Cdt-km.d.cts}

@@ -74,12 +74,16 @@ interface PromptRequest {
     maxTokens?: number;
     /**
      * Provider-specific request-body parameters merged verbatim into the
-     * outgoing request by OpenAI-compatible adapters. Opt-in escape hatch for
-     * fields the typed interface does not model — e.g. a vLLM/Mantle-served
-     * Kimi endpoint's `{chat_template_kwargs: {enable_thinking: false}}` to
-     * suppress chain-of-thought. Never sent unless set. Keys that collide with
-     * library-managed fields (model, messages, stream, tools, …) are ignored,
-     * so this can only add parameters, never override core request shape.
+     * outgoing request by OpenAI-compatible adapters and the Anthropic adapter
+     * (the Gemini adapters merge them into the generation config). Opt-in
+     * escape hatch for fields the typed interface does not model — e.g. a
+     * vLLM/Mantle-served Kimi endpoint's
+     * `{chat_template_kwargs: {enable_thinking: false}}` to suppress
+     * chain-of-thought, or Claude's `{output_config: {effort: 'low'}}` to cap
+     * how much it thinks before answering. Never sent unless set. Keys that
+     * collide with library-managed fields (model, messages, stream, tools, …)
+     * are ignored, so this can only add parameters, never override core
+     * request shape.
      */
     providerParams?: Record<string, unknown>;
     /** Abort the stream. Adapters MUST honor this and propagate to the vendor SDK. */
@@ -115,6 +119,16 @@ type LlmEvent = {
     id: string;
     name: string;
     arguments: unknown;
+    /**
+     * Opaque vendor-native state the adapter needs back when the call is
+     * written into history: pass the event to `appendAssistantToolCall`
+     * unchanged. Anthropic records the turn's content blocks here
+     * (`thinking`, `redacted_thinking`, text and `tool_use`, in order) and
+     * replays them, so the model keeps its reasoning after the tool result
+     * instead of working the turn out again. Absent unless the adapter
+     * set it.
+     */
+    vendorRaw?: unknown;
 } | {
     type: 'end';
     finishReason: FinishReason;

package/dist/{types-C_yrWq_0.d.ts → types-C3Cdt-km.d.ts}

@@ -74,12 +74,16 @@ interface PromptRequest {
     maxTokens?: number;
     /**
      * Provider-specific request-body parameters merged verbatim into the
-     * outgoing request by OpenAI-compatible adapters. Opt-in escape hatch for
-     * fields the typed interface does not model — e.g. a vLLM/Mantle-served
-     * Kimi endpoint's `{chat_template_kwargs: {enable_thinking: false}}` to
-     * suppress chain-of-thought. Never sent unless set. Keys that collide with
-     * library-managed fields (model, messages, stream, tools, …) are ignored,
-     * so this can only add parameters, never override core request shape.
+     * outgoing request by OpenAI-compatible adapters and the Anthropic adapter
+     * (the Gemini adapters merge them into the generation config). Opt-in
+     * escape hatch for fields the typed interface does not model — e.g. a
+     * vLLM/Mantle-served Kimi endpoint's
+     * `{chat_template_kwargs: {enable_thinking: false}}` to suppress
+     * chain-of-thought, or Claude's `{output_config: {effort: 'low'}}` to cap
+     * how much it thinks before answering. Never sent unless set. Keys that
+     * collide with library-managed fields (model, messages, stream, tools, …)
+     * are ignored, so this can only add parameters, never override core
+     * request shape.
      */
     providerParams?: Record<string, unknown>;
     /** Abort the stream. Adapters MUST honor this and propagate to the vendor SDK. */
@@ -115,6 +119,16 @@ type LlmEvent = {
     id: string;
     name: string;
     arguments: unknown;
+    /**
+     * Opaque vendor-native state the adapter needs back when the call is
+     * written into history: pass the event to `appendAssistantToolCall`
+     * unchanged. Anthropic records the turn's content blocks here
+     * (`thinking`, `redacted_thinking`, text and `tool_use`, in order) and
+     * replays them, so the model keeps its reasoning after the tool result
+     * instead of working the turn out again. Absent unless the adapter
+     * set it.
+     */
+    vendorRaw?: unknown;
 } | {
     type: 'end';
     finishReason: FinishReason;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jambonz/llm",
-  "version": "0.6.9",
+  "version": "0.6.11",
   "description": "Voice-oriented LLM adapter library — a uniform interface for streaming chat completions from multiple LLM vendors, with first-class support for tool calls, AbortSignal interruption, and lossless history round-tripping.",
   "type": "module",
   "main": "./dist/index.cjs",