@wix/evalforge-evaluator 0.114.0 → 0.116.0

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

@@ -13,6 +13,11 @@ export declare class ClaudeCodeAdapter implements AgentAdapter {
     readonly id = "claude-code";
     readonly name = "Claude Code";
     readonly supportedCommands: readonly [AgentRunCommand.CLAUDE];
+    /**
+     * Write infrastructure files (settings, MCPs, sub-agents, rules, skills)
+     * before the baseline snapshot is taken.
+     */
+    prepareEnvironment(context: AgentExecutionContext): Promise<void>;
     /**
      * Execute a skill using the Claude Code SDK.
      *

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

@@ -13,6 +13,12 @@ export interface TimestampedMessage {
     message: SDKMessage;
     receivedAt: Date;
 }
+/**
+ * Write all infrastructure files (settings, MCPs, sub-agents, rules, skills)
+ * to the working directory. Called by the adapter's `prepareEnvironment()` so
+ * that the orchestrator can take the baseline snapshot *after* infra is in place.
+ */
+export declare function prepareClaudeCodeEnvironment(cwd: string, skills: SkillWithLatestVersion[], options: Pick<ClaudeCodeExecutionOptions, 'mcps' | 'subAgents' | 'rules'>): Promise<void>;
 /**
  * Execute skills using the Claude Agent SDK.
  *

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

@@ -1,5 +1,13 @@
 import type { SkillWithLatestVersion, TestScenario, LLMTrace, ConversationMessage } from '@wix/evalforge-types';
 import type { OpenCodeExecutionOptions, OpenCodeExecutionResult } from './types.js';
+/**
+ * Write all infrastructure files (sub-agents, rules, skills)
+ * to the working directory. Called by the adapter's `prepareEnvironment()` so
+ * that the orchestrator can take the baseline snapshot *after* infra is in place.
+ *
+ * Note: MCPs for OpenCode are passed inline via config, not written to filesystem.
+ */
+export declare function prepareOpenCodeEnvironment(cwd: string, skills: SkillWithLatestVersion[], options: Pick<OpenCodeExecutionOptions, 'mcps' | 'subAgents' | 'rules'>): Promise<void>;
 /**
  * Execute skills using the OpenCode SDK.
  *

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

@@ -13,6 +13,7 @@ export declare class OpenCodeAdapter implements AgentAdapter {
     readonly id = "opencode";
     readonly name = "OpenCode";
     readonly supportedCommands: readonly [AgentRunCommand.OPENCODE];
+    prepareEnvironment(context: AgentExecutionContext): Promise<void>;
     execute(context: AgentExecutionContext): Promise<AgentExecutionResult>;
 }
 export declare const openCodeAdapter: OpenCodeAdapter;

package/build/types/run-scenario/file-diff.d.ts

@@ -5,6 +5,12 @@
  * when running against a template.
  */
 import type { DiffContent, TemplateFile } from '@wix/evalforge-types';
+/**
+ * Derive the set of infrastructure file paths by comparing a pre-prepare
+ * snapshot with the post-prepare (baseline) snapshot.
+ * Any file that was created or modified by `prepareEnvironment()` is infrastructure.
+ */
+export declare function deriveInfrastructurePaths(prePrep: FileSnapshot, postPrep: FileSnapshot): Set<string>;
 /**
  * A snapshot of file contents in a directory.
  * Maps relative file paths to their contents.
@@ -26,14 +32,16 @@ export declare function snapshotDirectory(dir: string, baseDir?: string): FileSn
  *
  * @param before - Snapshot before execution
  * @param after - Snapshot after execution
+ * @param infrastructurePaths - Optional set of paths known to be infrastructure files
  * @returns Array of DiffContent for files that were created, modified, or renamed
  */
-export declare function diffSnapshots(before: FileSnapshot, after: FileSnapshot): DiffContent[];
+export declare function diffSnapshots(before: FileSnapshot, after: FileSnapshot, infrastructurePaths?: Set<string>): DiffContent[];
 /**
  * Extract template files with their status from before/after snapshots.
  *
  * @param before - Snapshot before execution
  * @param after - Snapshot after execution
+ * @param infrastructurePaths - Optional set of paths known to be infrastructure files
  * @returns Array of TemplateFile with status indicators
  */
-export declare function extractTemplateFiles(before: FileSnapshot, after: FileSnapshot): TemplateFile[];
+export declare function extractTemplateFiles(before: FileSnapshot, after: FileSnapshot, infrastructurePaths?: Set<string>): TemplateFile[];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wix/evalforge-evaluator",
-  "version": "0.114.0",
+  "version": "0.116.0",
   "description": "EvalForge Evaluator",
   "bin": "./build/index.js",
   "files": [
@@ -22,9 +22,9 @@
     "@anthropic-ai/claude-agent-sdk": "^0.2.63",
     "@anthropic-ai/claude-code": "^2.1.63",
     "@opencode-ai/sdk": "^1.2.15",
-    "@wix/eval-assertions": "0.27.0",
-    "@wix/evalforge-github-client": "0.32.0",
-    "@wix/evalforge-types": "0.57.0",
+    "@wix/eval-assertions": "0.28.0",
+    "@wix/evalforge-github-client": "0.33.0",
+    "@wix/evalforge-types": "0.58.0",
     "ai": "^6.0.107",
     "diff": "^7.0.0",
     "tar": "^7.5.3",
@@ -63,5 +63,5 @@
       "artifactId": "evalforge-evaluator"
     }
   },
-  "falconPackageHash": "7580e561b029720761957e0a2ecad733cd2bc610cfe82d11e3c8060a"
+  "falconPackageHash": "9c7374a0596e12fcb8bdb6e1c2fd18bb9160ac34ed39cc196975dd98"
 }