@compilr-dev/agents 0.4.1 → 0.5.0

package/dist/agent.js

@@ -2499,6 +2499,13 @@ export class Agent {
                     // Capture usage data from done chunk
                     usage = chunk.usage;
                     model = chunk.model;
+                    // Handle special stop reasons (Claude 4.5+)
+                    if (chunk.stopReason === 'refusal') {
+                        text += '\n\n[Model refused this request]';
+                    }
+                    else if (chunk.stopReason === 'context_window_exceeded') {
+                        text += '\n\n[Response truncated — context window limit reached]';
+                    }
                     break;
             }
         }

package/dist/providers/claude.d.ts

@@ -87,6 +87,10 @@ export declare class ClaudeProvider implements LLMProvider {
      * Count tokens in messages using tiktoken (cl100k_base encoding)
      */
     countTokens(messages: Message[]): Promise<number>;
+    /**
+     * Check if a model is Claude 4+ (token-efficient tools are built-in, no header needed).
+     */
+    private isClaude4Plus;
     /**
      * Build request options with optional abort signal and beta headers.
      * Combines multiple beta features (comma-separated per Anthropic API spec).
@@ -105,7 +109,10 @@ export declare class ClaudeProvider implements LLMProvider {
      */
     private convertTools;
     /**
-     * Convert thinking config to Anthropic API format
+     * Convert thinking config to Anthropic API format.
+     *
+     * Supports both legacy (enabled + budget_tokens) and adaptive (Claude 4.6+) modes.
+     * Adaptive mode uses output_config.effort instead of budget_tokens.
      */
     private convertThinking;
     /**

package/dist/providers/claude.js

@@ -79,13 +79,16 @@ export class ClaudeProvider {
                 temperature: options?.temperature,
                 stop_sequences: options?.stopSequences,
             };
-            // Add thinking if enabled (Claude-specific)
+            // Add thinking + effort if configured (Claude-specific)
             // Note: Extended thinking types not yet in SDK, using Object.assign
             if (thinking) {
-                Object.assign(params, { thinking });
+                Object.assign(params, { thinking: thinking.thinking });
+                if (thinking.outputConfig) {
+                    Object.assign(params, { output_config: thinking.outputConfig });
+                }
             }
             // Pass abort signal and optional beta header to SDK
-            const requestOptions = this.buildRequestOptions(options?.signal, tools.length > 0);
+            const requestOptions = this.buildRequestOptions(options?.signal, tools.length > 0, options?.model);
             const stream = this.client.messages.stream(params, requestOptions);
             const model = options?.model ?? this.defaultModel;
             let currentToolId = '';
@@ -118,9 +121,16 @@ export class ClaudeProvider {
             // Access optional cache token fields via type coercion
             // These fields are present in newer SDK versions but not in the type definitions
             const usageWithCache = usage;
+            // Map Anthropic stop_reason to our stopReason
+            // Cast to string — SDK types may not include newer stop reasons (refusal, model_context_window_exceeded)
+            const rawStopReason = finalMessage.stop_reason;
+            const stopReason = rawStopReason === 'model_context_window_exceeded'
+                ? 'context_window_exceeded'
+                : (rawStopReason ?? 'end_turn');
             yield {
                 type: 'done',
                 model,
+                stopReason,
                 usage: {
                     inputTokens: usage.input_tokens,
                     outputTokens: usage.output_tokens,
@@ -140,13 +150,22 @@ export class ClaudeProvider {
     countTokens(messages) {
         return Promise.resolve(countMessageTokens(messages));
     }
+    /**
+     * Check if a model is Claude 4+ (token-efficient tools are built-in, no header needed).
+     */
+    isClaude4Plus(model) {
+        // Claude 4+ model IDs: claude-opus-4-*, claude-sonnet-4-*, claude-haiku-4-*
+        return /^claude-(opus|sonnet|haiku)-4/.test(model);
+    }
     /**
      * Build request options with optional abort signal and beta headers.
      * Combines multiple beta features (comma-separated per Anthropic API spec).
      */
-    buildRequestOptions(signal, hasTools) {
+    buildRequestOptions(signal, hasTools, model) {
+        const resolvedModel = model ?? this.defaultModel;
         const betas = [];
-        if (this.enableTokenEfficientTools && hasTools) {
+        // token-efficient-tools is built-in for Claude 4+ — only send for older models
+        if (this.enableTokenEfficientTools && hasTools && !this.isClaude4Plus(resolvedModel)) {
             betas.push('token-efficient-tools-2025-02-19');
         }
         if (this.enableExtendedContext) {
@@ -235,19 +254,30 @@ export class ClaudeProvider {
         }));
     }
     /**
-     * Convert thinking config to Anthropic API format
+     * Convert thinking config to Anthropic API format.
+     *
+     * Supports both legacy (enabled + budget_tokens) and adaptive (Claude 4.6+) modes.
+     * Adaptive mode uses output_config.effort instead of budget_tokens.
      */
     convertThinking(thinking) {
         if (!thinking || thinking.type === 'disabled') {
             return undefined;
         }
-        // Validate budget_tokens minimum (1024)
-        if (thinking.budgetTokens < 1024) {
-            throw new ProviderError(`Extended thinking budget_tokens must be at least 1024, got ${String(thinking.budgetTokens)}`, 'claude');
+        if (thinking.type === 'adaptive') {
+            return {
+                thinking: { type: 'adaptive' },
+                outputConfig: thinking.effort ? { effort: thinking.effort } : undefined,
+            };
+        }
+        // Legacy enabled mode (deprecated on Claude 4.6, still works on older)
+        if (!thinking.budgetTokens || thinking.budgetTokens < 1024) {
+            throw new ProviderError(`Extended thinking budget_tokens must be at least 1024, got ${String(thinking.budgetTokens ?? 0)}`, 'claude');
         }
         return {
-            type: thinking.type,
-            budget_tokens: thinking.budgetTokens,
+            thinking: {
+                type: 'enabled',
+                budget_tokens: thinking.budgetTokens,
+            },
         };
     }
     /**

package/dist/providers/types.d.ts

@@ -110,6 +110,15 @@ export interface StreamChunk {
      * Model that generated this response (only present on 'done' chunks)
      */
     model?: string;
+    /**
+     * Stop reason (only present on 'done' chunks).
+     * - 'end_turn': Normal completion
+     * - 'max_tokens': Hit max_tokens limit
+     * - 'refusal': Model refused the request (Claude 4.5+)
+     * - 'context_window_exceeded': Hit context window limit (Claude 4.5+)
+     * - 'tool_use': Model wants to call a tool
+     */
+    stopReason?: string;
 }
 /**
  * Extended thinking configuration
@@ -118,13 +127,26 @@ export interface StreamChunk {
  */
 export interface ThinkingConfig {
     /**
-     * Enable or disable extended thinking
+     * Thinking mode:
+     * - 'enabled': Manual budget (legacy, deprecated on Claude 4.6)
+     * - 'adaptive': Model decides when/how much to think (Claude 4.6+)
+     * - 'disabled': No thinking
+     */
+    type: 'enabled' | 'adaptive' | 'disabled';
+    /**
+     * Token budget for thinking (minimum 1024, must be less than maxTokens).
+     * Only used with type: 'enabled'. Ignored for 'adaptive'.
      */
-    type: 'enabled' | 'disabled';
+    budgetTokens?: number;
     /**
-     * Token budget for thinking (minimum 1024, must be less than maxTokens)
+     * Effort level for adaptive thinking (Claude 4.6+).
+     * Controls how much the model thinks before responding.
+     * - 'low': Minimal thinking, fastest response
+     * - 'medium': Balanced
+     * - 'high': Deep thinking, most thorough (Sonnet 4.6 default)
+     * Only used with type: 'adaptive'. Ignored for 'enabled'.
      */
-    budgetTokens: number;
+    effort?: 'low' | 'medium' | 'high';
 }
 /**
  * Options for chat requests

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@compilr-dev/agents",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "Lightweight multi-LLM agent library for building CLI AI assistants",
   "type": "module",
   "main": "dist/index.js",