@wix/evalforge-evaluator 0.139.0 → 0.140.0

package/build/types/run-scenario/agents/claude-code/execute.d.ts

@@ -1,11 +1,11 @@
-import type { SkillWithLatestVersion, TestScenario, LLMTrace, ConversationMessage } from '@wix/evalforge-types';
+import type { SkillWithLatestVersion, TestScenario, LLMTrace, ConversationMessage, TriggerPromptImage } from '@wix/evalforge-types';
 import type { ClaudeCodeExecutionOptions, ClaudeCodeExecutionResult } from './types.js';
 /**
  * Import SDK types directly from Claude Agent SDK.
  * Type-only imports are erased at compile time - zero runtime overhead.
  * The SDK is still dynamically imported at runtime in executeWithClaudeCode().
  */
-import type { SDKMessage } from '@anthropic-ai/claude-agent-sdk' with { 'resolution-mode': 'import' };
+import type { SDKUserMessage, SDKMessage } from '@anthropic-ai/claude-agent-sdk' with { 'resolution-mode': 'import' };
 /**
  * Message with timestamp — tracks when each message was received.
  */
@@ -13,6 +13,16 @@ export interface TimestampedMessage {
     message: SDKMessage;
     receivedAt: Date;
 }
+/**
+ * Build an async iterable that yields a user message with images.
+ * When images are attached to the trigger prompt, this allows multimodal input
+ * to the Claude Agent SDK via content blocks (text + image).
+ *
+ * @param triggerPrompt - The text prompt to send
+ * @param images - Array of images with base64 data and metadata
+ * @returns An async iterable yielding a single SDKUserMessage
+ */
+export declare function buildPromptStream(triggerPrompt: string, images: TriggerPromptImage[]): AsyncIterable<SDKUserMessage>;
 /**
  * Write all infrastructure files (settings, MCPs, sub-agents, rules, skills)
  * to the working directory. Called by the adapter's `prepareEnvironment()` so

package/build/types/run-scenario/agents/opencode/execute.d.ts

@@ -1,5 +1,17 @@
-import type { SkillWithLatestVersion, TestScenario, LLMTrace, ConversationMessage } from '@wix/evalforge-types';
+import type { SkillWithLatestVersion, TestScenario, LLMTrace, ConversationMessage, TriggerPromptImage } from '@wix/evalforge-types';
 import type { OpenCodeExecutionOptions, OpenCodeExecutionResult } from './types.js';
+/**
+ * Write trigger prompt images to disk so they can be passed via --file flags.
+ *
+ * Note: OpenCode CLI's --file flag currently has a known MIME-type detection bug
+ * (hardcodes text/plain for all files). Images are still written to the cwd so the
+ * agent can access them via its tools even if the CLI doesn't render them visually.
+ *
+ * @param cwd - Working directory where images will be written
+ * @param images - Array of images with base64 data and metadata
+ * @returns Array of absolute file paths for the written images
+ */
+export declare function writePromptImages(cwd: string, images: TriggerPromptImage[]): Promise<string[]>;
 /**
  * Write all infrastructure files (sub-agents, rules, skills)
  * to the working directory. Called by the adapter's `prepareEnvironment()` so

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wix/evalforge-evaluator",
-  "version": "0.139.0",
+  "version": "0.140.0",
   "description": "EvalForge Evaluator",
   "bin": "./build/index.js",
   "files": [
@@ -21,9 +21,9 @@
     "@ai-sdk/openai": "^3.0.39",
     "@anthropic-ai/claude-agent-sdk": "^0.2.63",
     "@anthropic-ai/claude-code": "^2.1.63",
-    "@wix/eval-assertions": "0.42.0",
-    "@wix/evalforge-github-client": "0.46.0",
-    "@wix/evalforge-types": "0.71.0",
+    "@wix/eval-assertions": "0.43.0",
+    "@wix/evalforge-github-client": "0.47.0",
+    "@wix/evalforge-types": "0.72.0",
     "ai": "^6.0.107",
     "diff": "^7.0.0",
     "tar": "^7.5.3",
@@ -62,5 +62,5 @@
       "artifactId": "evalforge-evaluator"
     }
   },
-  "falconPackageHash": "5bb57c48b8bf1e552339edfe3952c4156ed91c946c2701fad8e2778f"
+  "falconPackageHash": "ff3677bd864b44dbb26d0cf845d0e17f4039fe8031366a11a13e2cf1"
 }