@iqai/adk 0.5.6 → 0.5.9

package/CHANGELOG.md

@@ -1,5 +1,50 @@
 # @iqai/adk
 
+## 0.5.9
+
+### Patch Changes
+
+- 27d6bd9: Refactored MCP sampling parameters to align with the official protocol specification by moving from direct model assertions to the `modelPreferences` hint system.
+
+  - **Protocol Alignment**
+    Removed the invalid `mcpParams.model` type assertion, as the MCP spec does not define a top-level string for the model in sampling requests.
+
+  - **Preference Logic**
+    Implemented support for `modelPreferences.hints`, allowing the server to interpret model suggestions via the `name` property within the hints array.
+
+  - **Resilience**
+    Added optional chaining across the parameter parsing logic to prevent runtime errors when `modelPreferences` or `hints` are undefined.
+
+  - **Fallback Strategy**
+    Established `gemini-2.0-flash` as the default model if the client provides no specific hints or valid preferences.
+
+## 0.5.8
+
+### Patch Changes
+
+- d92892c: - Added aggregated token usage tracking (input, output, total) across agents, nested agents, and LLM generations.
+  - Introduced model usage tracking per agent and invocation for improved model attribution.
+  - Embedded aggregated token and model usage data into Langfuse agent spans and root trace metadata.
+  - Ensured compatibility with nested agents without modifying the existing span hierarchy.
+  - Implemented cleanup of internal token and model tracking maps after each invocation to prevent memory leaks.
+  - Improved event naming with descriptive suffixes (`.function_call`, `.final_response`, `.event`).
+  - Included `finishReason` in event metadata for enhanced execution context and observability.
+- 0ada268: Added full plugin support across the agent lifecycle, enabling interception and extension of agent behavior during execution, model calls, and tool invocations.
+
+  - Integrated plugin manager into `BaseAgent`, giving plugins priority over canonical callbacks.
+  - Added `plugins` configuration to `AgentBuilder` and introduced `withPlugins()` API.
+  - Updated `LlmAgent` and `BaseAgent` flows to run plugin lifecycle hooks (`before/after agent`, `before/after model`, `before/after tool`, error hooks, etc.).
+  - Plugins can now override or modify behavior at multiple stages.
+
+  - Fully backward compatible (optional chaining, canonical callbacks still work).
+  - Enables custom logging, monitoring, request rewriting, caching, rate limiting, and more via plugins.
+
+## 0.5.7
+
+### Patch Changes
+
+- b938be4: fix: underlying type errors with Model Context Protocol versioning
+
 ## 0.5.6
 
 ### Patch Changes

package/README.md

@@ -88,7 +88,7 @@ import dotenv from 'dotenv';
 dotenv.config();
 
 // Instantiate the agent
-const myAgent = new Agent({
+const myAgent = new LlmAgent({
   name: "simple_query_assistant",
   model: "gemini-2.5-flash", // Or "gpt-4-turbo", "claude-3-opus"
   description: "A basic assistant to answer questions.",
@@ -229,7 +229,7 @@ class CalculatorTool extends BaseTool {
 }
 
 // Create an agent equipped with the calculator tool
-const mathAgent = new Agent({
+const mathAgent = new LlmAgent({
   name: "math_assistant_agent",
   model: "gpt-4-turbo", // Choose a model proficient with tool usage
   instructions:
@@ -276,7 +276,7 @@ const parallelAnalysis = await AgentBuilder
 ```typescript
 import { Agent, InMemorySessionService } from '@iqai/adk';
 
-const agent = new Agent({
+const agent = new LlmAgent({
   name: "persistent_assistant",
   model: "gpt-4-turbo",
   sessionService: new InMemorySessionService(),