fix(workspace): default hook cwd to workspaceFileDir for file-referenced templates

packages/core/src/evaluation/orchestrator.ts

@@ -962,6 +962,7 @@ export async function runEvaluation(
         testId: '__before_all__',
         evalRunId,
         evalDir,
+        workspaceFileDir: suiteWorkspace?.workspaceFileDir,
       };
       try {
         beforeAllOutput = await executeWorkspaceScript(
@@ -988,6 +989,7 @@ export async function runEvaluation(
           testId: '__before_all__',
           evalRunId,
           evalDir,
+          workspaceFileDir: suiteWorkspace?.workspaceFileDir,
         };
         try {
           const output = await executeWorkspaceScript(
@@ -1408,6 +1410,7 @@ export async function runEvaluation(
           testId: '__after_all__',
           evalRunId,
           evalDir,
+          workspaceFileDir: suiteWorkspace?.workspaceFileDir,
         };
         try {
           const afterAllOutput = await executeWorkspaceScript(
@@ -1859,6 +1862,7 @@ export async function runEvalCase(options: RunEvalCaseOptions): Promise<Evaluati
         caseInput: evalCase.question,
         caseMetadata: evalCase.metadata,
         evalDir,
+        workspaceFileDir: evalCase.workspace?.workspaceFileDir,
       };
       try {
         beforeAllOutput = await executeWorkspaceScript(
@@ -1938,6 +1942,7 @@ export async function runEvalCase(options: RunEvalCaseOptions): Promise<Evaluati
       caseInput: evalCase.question,
       caseMetadata: evalCase.metadata,
       evalDir,
+      workspaceFileDir: evalCase.workspace?.workspaceFileDir,
     };
     try {
       beforeEachOutput = await executeWorkspaceScript(
@@ -2195,6 +2200,7 @@ export async function runEvalCase(options: RunEvalCaseOptions): Promise<Evaluati
       caseInput: evalCase.question,
       caseMetadata: evalCase.metadata,
       evalDir,
+      workspaceFileDir: evalCase.workspace?.workspaceFileDir,
     };
     try {
       afterEachOutput = await executeWorkspaceScript(

packages/core/src/evaluation/types.ts

@@ -321,6 +321,10 @@ export type WorkspaceConfig = {
   readonly path?: string;
   /** Docker-based workspace: run grader commands inside a container */
   readonly docker?: DockerWorkspaceConfig;
+  /** Directory containing the workspace file when workspace is a file reference.
+   *  Used as default cwd for hook commands so that file-referenced templates resolve
+   *  relative paths from their own directory, not the eval file's directory. */
+  readonly workspaceFileDir?: string;
 };
 
 export type CodeEvaluatorConfig = {

packages/core/src/evaluation/workspace/script-executor.ts

@@ -10,8 +10,12 @@ export interface ScriptExecutionContext {
   readonly evalRunId: string;
   readonly caseInput?: string;
   readonly caseMetadata?: Record<string, unknown>;
-  /** Directory containing the eval YAML file. Used as default cwd. */
+  /** Directory containing the eval YAML file. Used as fallback cwd. */
   readonly evalDir?: string;
+  /** Directory containing the workspace file (when workspace is a file reference).
+   *  Takes priority over evalDir as default cwd so that file-referenced templates
+   *  resolve relative paths from their own directory. */
+  readonly workspaceFileDir?: string;
 }
 
 export type ScriptFailureMode = 'fatal' | 'warn';
@@ -57,7 +61,7 @@ export async function executeWorkspaceScript(
   });
 
   const timeoutMs = config.timeout_ms ?? (failureMode === 'fatal' ? 60000 : 30000);
-  const cwd = config.cwd ?? context.evalDir;
+  const cwd = config.cwd ?? context.workspaceFileDir ?? context.evalDir;
 
   // Support both command (canonical) and script (deprecated alias)
   if (config.script !== undefined && config.command === undefined) {

packages/core/src/evaluation/yaml-parser.ts

@@ -751,7 +751,7 @@ async function resolveWorkspaceConfig(
     const workspaceFileDir = path.dirname(workspaceFilePath);
     const resolvedWorkspace = parseWorkspaceConfig(parsed, workspaceFileDir);
     if (resolvedWorkspace) {
-      return resolvedWorkspace;
+      return { ...resolvedWorkspace, workspaceFileDir };
     }
 
     const parsedObject = parsed as Record<string, unknown>;
@@ -882,6 +882,7 @@ function mergeWorkspaceConfigs(
     mode: caseLevel.mode ?? suiteLevel.mode,
     path: caseLevel.path ?? suiteLevel.path,
     docker: caseLevel.docker ?? suiteLevel.docker,
+    workspaceFileDir: caseLevel.workspaceFileDir ?? suiteLevel.workspaceFileDir,
   };
 }
 

packages/core/test/evaluation/workspace-config-parsing.test.ts

@@ -557,6 +557,62 @@ tests:
       expect(cases[0].workspace?.template).toBe(path.join(wsDir, 'my-template'));
       // cwd resolved relative to workspace file dir
       expect(cases[0].workspace?.hooks?.before_all?.cwd).toBe(path.join(wsDir, 'scripts'));
+      // workspaceFileDir is set to the workspace file's directory
+      expect(cases[0].workspace?.workspaceFileDir).toBe(wsDir);
+    });
+
+    it('should set workspaceFileDir when workspace is a file reference', async () => {
+      const wsDir = path.join(testDir, 'wsfiledir-test');
+      await mkdir(wsDir, { recursive: true });
+
+      const workspaceFile = path.join(wsDir, 'workspace.yaml');
+      await writeFile(
+        workspaceFile,
+        `
+hooks:
+  before_all:
+    command: ["echo", "hello"]
+`,
+      );
+
+      const evalFile = path.join(testDir, 'wsfiledir-eval.yaml');
+      await writeFile(
+        evalFile,
+        `
+workspace: ./wsfiledir-test/workspace.yaml
+
+tests:
+  - id: wsfiledir-test-1
+    input: "Do something"
+    criteria: "Should work"
+`,
+      );
+
+      const cases = await loadTests(evalFile, testDir);
+      expect(cases).toHaveLength(1);
+      expect(cases[0].workspace?.workspaceFileDir).toBe(wsDir);
+    });
+
+    it('should not set workspaceFileDir for inline workspace config', async () => {
+      const evalFile = path.join(testDir, 'inline-workspace.yaml');
+      await writeFile(
+        evalFile,
+        `
+workspace:
+  hooks:
+    before_all:
+      command: ["echo", "hello"]
+
+tests:
+  - id: inline-test-1
+    input: "Do something"
+    criteria: "Should work"
+`,
+      );
+
+      const cases = await loadTests(evalFile, testDir);
+      expect(cases).toHaveLength(1);
+      expect(cases[0].workspace?.workspaceFileDir).toBeUndefined();
     });
 
     it('should throw a clear error when workspace file is not found', async () => {

packages/core/test/evaluation/workspace/script-executor.test.ts

@@ -344,4 +344,69 @@ process.stdout.write(JSON.stringify(args));
       await rm(explicitDir, { recursive: true, force: true });
     }
   });
+
+  it('defaults cwd to workspaceFileDir over evalDir when workspace is a file reference', async () => {
+    const evalDir = path.join(tmpdir(), `agentv-evaldir-${randomUUID()}`);
+    const workspaceFileDir = path.join(tmpdir(), `agentv-wsfiledir-${randomUUID()}`);
+    await mkdir(evalDir, { recursive: true });
+    await mkdir(workspaceFileDir, { recursive: true });
+
+    try {
+      const cwdScript = path.join(testDir, 'print-cwd3.js');
+      await writeFile(cwdScript, 'process.stdout.write(process.cwd());');
+
+      const config: WorkspaceScriptConfig = {
+        command: ['node', cwdScript],
+        // No cwd — should default to workspaceFileDir, not evalDir
+      };
+
+      const context: ScriptExecutionContext = {
+        workspacePath: '/tmp/workspace',
+        testId: 'wsfiledir-default-test',
+        evalRunId: 'run-cwd-3',
+        evalDir,
+        workspaceFileDir,
+      };
+
+      const output = await executeWorkspaceScript(config, context, 'fatal');
+      expect(output).toBe(workspaceFileDir);
+    } finally {
+      await rm(evalDir, { recursive: true, force: true });
+      await rm(workspaceFileDir, { recursive: true, force: true });
+    }
+  });
+
+  it('uses explicit cwd over workspaceFileDir', async () => {
+    const evalDir = path.join(tmpdir(), `agentv-evaldir-${randomUUID()}`);
+    const workspaceFileDir = path.join(tmpdir(), `agentv-wsfiledir-${randomUUID()}`);
+    const explicitDir = path.join(tmpdir(), `agentv-explicit-${randomUUID()}`);
+    await mkdir(evalDir, { recursive: true });
+    await mkdir(workspaceFileDir, { recursive: true });
+    await mkdir(explicitDir, { recursive: true });
+
+    try {
+      const cwdScript = path.join(testDir, 'print-cwd4.js');
+      await writeFile(cwdScript, 'process.stdout.write(process.cwd());');
+
+      const config: WorkspaceScriptConfig = {
+        command: ['node', cwdScript],
+        cwd: explicitDir, // Explicit cwd should override both workspaceFileDir and evalDir
+      };
+
+      const context: ScriptExecutionContext = {
+        workspacePath: '/tmp/workspace',
+        testId: 'explicit-over-wsfiledir-test',
+        evalRunId: 'run-cwd-4',
+        evalDir,
+        workspaceFileDir,
+      };
+
+      const output = await executeWorkspaceScript(config, context, 'fatal');
+      expect(output).toBe(explicitDir);
+    } finally {
+      await rm(evalDir, { recursive: true, force: true });
+      await rm(workspaceFileDir, { recursive: true, force: true });
+      await rm(explicitDir, { recursive: true, force: true });
+    }
+  });
 });