agentv 4.40.1 → 4.41.0-next.1

package/dist/{chunk-BLXYBUU4.js → chunk-ENHX2CCS.js}

@@ -493,8 +493,8 @@ function getErrorMap() {
 
 // ../../node_modules/.bun/zod@3.25.76/node_modules/zod/v3/helpers/parseUtil.js
 var makeIssue = (params) => {
-  const { data, path: path50, errorMaps, issueData } = params;
-  const fullPath = [...path50, ...issueData.path || []];
+  const { data, path: path51, errorMaps, issueData } = params;
+  const fullPath = [...path51, ...issueData.path || []];
   const fullIssue = {
     ...issueData,
     path: fullPath
@@ -610,11 +610,11 @@ var errorUtil;
 
 // ../../node_modules/.bun/zod@3.25.76/node_modules/zod/v3/types.js
 var ParseInputLazyPath = class {
-  constructor(parent, value, path50, key) {
+  constructor(parent, value, path51, key) {
     this._cachedPath = [];
     this.parent = parent;
     this.data = value;
-    this._path = path50;
+    this._path = path51;
     this._key = key;
   }
   get path() {
@@ -4056,7 +4056,7 @@ var coerce = {
 };
 var NEVER = INVALID;
 
-// ../../packages/core/dist/chunk-5JNFEE7J.js
+// ../../packages/core/dist/chunk-3EAL7M5J.js
 import { parse } from "yaml";
 import os from "node:os";
 import path from "node:path";
@@ -5252,17 +5252,22 @@ function resolveCopilotFlatProviderConfig(target, env) {
       optionalEnv: true
     }
   );
-  const wireApi = resolveOptionalString(target.wire_api, env, `${target.name} copilot wire API`, {
-    allowLiteral: true,
-    optionalEnv: true
-  });
+  const apiFormat = resolveOptionalString(
+    target.api_format,
+    env,
+    `${target.name} copilot API format`,
+    {
+      allowLiteral: true,
+      optionalEnv: true
+    }
+  );
   return {
     ...type ? { type } : {},
     baseUrl,
     ...apiKey ? { apiKey } : {},
     ...bearerToken ? { bearerToken } : {},
     ...apiVersion ? { apiVersion } : {},
-    ...wireApi ? { wireApi } : {}
+    ...apiFormat ? { wireApi: apiFormat } : {}
   };
 }
 function resolveCopilotCliConfig(target, env, _evalFilePath) {
@@ -6151,21 +6156,17 @@ async function expandFileReferences(tests, evalFileDir) {
   return expanded;
 }
 
-// ../../packages/core/dist/chunk-M6LF2BEU.js
-import path49 from "node:path";
+// ../../packages/core/dist/chunk-REU6TJT4.js
+import path50 from "node:path";
 import { pathToFileURL as pathToFileURL2 } from "node:url";
 import { existsSync as existsSync7 } from "node:fs";
-import path48 from "node:path";
+import path49 from "node:path";
 import micromatch4 from "micromatch";
 import { mkdir, readFile as readFile3, writeFile } from "node:fs/promises";
 import path5 from "node:path";
-import { execFile as execFile2 } from "node:child_process";
 import { createHash as createHash5, randomUUID as randomUUID10 } from "node:crypto";
-import { existsSync as existsSync6 } from "node:fs";
-import { copyFile as copyFile2, mkdir as mkdir17, readdir as readdir8, stat as stat9 } from "node:fs/promises";
-import path46 from "node:path";
+import path47 from "node:path";
 import { fileURLToPath as fileURLToPath5 } from "node:url";
-import { promisify as promisify6 } from "node:util";
 import micromatch3 from "micromatch";
 import { mkdtemp, rm, writeFile as writeFile2 } from "node:fs/promises";
 import { tmpdir } from "node:os";
@@ -6883,10 +6884,10 @@ function assignProp(target, prop, value) {
     configurable: true
   });
 }
-function getElementAtPath(obj, path50) {
-  if (!path50)
+function getElementAtPath(obj, path51) {
+  if (!path51)
     return obj;
-  return path50.reduce((acc, key) => acc?.[key], obj);
+  return path51.reduce((acc, key) => acc?.[key], obj);
 }
 function promiseAllObject(promisesObj) {
   const keys = Object.keys(promisesObj);
@@ -7206,11 +7207,11 @@ function aborted(x, startIndex = 0) {
   }
   return false;
 }
-function prefixIssues(path50, issues) {
+function prefixIssues(path51, issues) {
   return issues.map((iss) => {
     var _a;
     (_a = iss).path ?? (_a.path = []);
-    iss.path.unshift(path50);
+    iss.path.unshift(path51);
     return iss;
   });
 }
@@ -7347,7 +7348,7 @@ function treeifyError(error40, _mapper) {
     return issue2.message;
   };
   const result = { errors: [] };
-  const processError = (error41, path50 = []) => {
+  const processError = (error41, path51 = []) => {
     var _a, _b;
     for (const issue2 of error41.issues) {
       if (issue2.code === "invalid_union" && issue2.errors.length) {
@@ -7357,7 +7358,7 @@ function treeifyError(error40, _mapper) {
       } else if (issue2.code === "invalid_element") {
         processError({ issues: issue2.issues }, issue2.path);
       } else {
-        const fullpath = [...path50, ...issue2.path];
+        const fullpath = [...path51, ...issue2.path];
         if (fullpath.length === 0) {
           result.errors.push(mapper(issue2));
           continue;
@@ -7387,9 +7388,9 @@ function treeifyError(error40, _mapper) {
   processError(error40);
   return result;
 }
-function toDotPath(path50) {
+function toDotPath(path51) {
   const segs = [];
-  for (const seg of path50) {
+  for (const seg of path51) {
     if (typeof seg === "number")
       segs.push(`[${seg}]`);
     else if (typeof seg === "symbol")
@@ -18819,7 +18820,7 @@ var RequestError = class _RequestError extends Error {
   }
 };
 
-// ../../packages/core/dist/chunk-M6LF2BEU.js
+// ../../packages/core/dist/chunk-REU6TJT4.js
 import { exec as execCallback } from "node:child_process";
 import { readdirSync, statSync } from "node:fs";
 import { readFile as readFile32, readdir as readdir2, stat as stat2 } from "node:fs/promises";
@@ -18894,6 +18895,11 @@ import path33 from "node:path";
 import fg3 from "fast-glob";
 import { cp, mkdir as mkdir14, readdir as readdir5, rm as rm4, stat as stat6 } from "node:fs/promises";
 import path34 from "node:path";
+import { execFile as execFile2 } from "node:child_process";
+import { existsSync as existsSync6 } from "node:fs";
+import { copyFile as copyFile2, mkdir as mkdir17, readdir as readdir8, stat as stat8 } from "node:fs/promises";
+import path39 from "node:path";
+import { promisify as promisify6 } from "node:util";
 import { createHash as createHash3 } from "node:crypto";
 import { existsSync as existsSync3 } from "node:fs";
 import { cp as cp2, mkdir as mkdir15, readFile as readFile11, readdir as readdir6, rm as rm5, unlink, writeFile as writeFile9 } from "node:fs/promises";
@@ -18909,28 +18915,28 @@ import path36 from "node:path";
 import { stringify as stringifyYaml } from "yaml";
 import { readdir as readdir7, stat as stat7 } from "node:fs/promises";
 import path38 from "node:path";
-import { readFile as readFile18, stat as stat8 } from "node:fs/promises";
-import path45 from "node:path";
+import { readFile as readFile18, stat as stat9 } from "node:fs/promises";
+import path46 from "node:path";
 import micromatch2 from "micromatch";
 import { stringify as stringifyYaml2 } from "yaml";
 import { readFile as readFile12 } from "node:fs/promises";
-import path39 from "node:path";
+import path40 from "node:path";
 import { readFile as readFile13 } from "node:fs/promises";
-import path41 from "node:path";
+import path422 from "node:path";
 import { constants as constants4 } from "node:fs";
 import { access as access4 } from "node:fs/promises";
-import path40 from "node:path";
+import path41 from "node:path";
 import { fileURLToPath as fileURLToPath4 } from "node:url";
 import { readFile as readFile15 } from "node:fs/promises";
-import path422 from "node:path";
+import path43 from "node:path";
 import { readFile as readFile14 } from "node:fs/promises";
 import { readFile as readFile17 } from "node:fs/promises";
-import path44 from "node:path";
+import path45 from "node:path";
 import micromatch from "micromatch";
 import { readFile as readFile16 } from "node:fs/promises";
-import path43 from "node:path";
+import path44 from "node:path";
 import { mkdir as mkdir18, readFile as readFile20, writeFile as writeFile10 } from "node:fs/promises";
-import path47 from "node:path";
+import path48 from "node:path";
 import { readFile as readFile19 } from "node:fs/promises";
 var DEFAULT_CACHE_PATH = ".agentv/cache";
 var ResponseCache = class {
@@ -22462,115 +22468,115 @@ var FieldAccuracyGrader = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path50, match, required: required2 = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path50);
-    const expectedValue = resolvePath(expectedData, path50);
+    const { path: path51, match, required: required2 = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path51);
+    const expectedValue = resolvePath(expectedData, path51);
     if (expectedValue === void 0) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path50}: no expected value`
+        message: `${path51}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required2) {
         return {
-          path: path50,
+          path: path51,
           score: 0,
           weight,
           hit: false,
-          message: `${path50} (required, missing)`
+          message: `${path51} (required, missing)`
         };
       }
       return {
-        path: path50,
+        path: path51,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path50}: optional field missing`
+        message: `${path51}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path50, candidateValue, expectedValue, weight);
+        return this.compareExact(path51, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path50,
+          path51,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path50, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path51, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path50,
+          path: path51,
           score: 0,
           weight,
           hit: false,
-          message: `${path50}: unknown match type "${match}"`
+          message: `${path51}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path50, candidateValue, expectedValue, weight) {
+  compareExact(path51, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: path50
+        message: path51
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path51} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (value mismatch)`
+      message: `${path51} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path50, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path51, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber(candidateValue);
     const expectedNum = toNumber(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (non-numeric value)`
+        message: `${path51} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (invalid numeric value)`
+        message: `${path51} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -22583,61 +22589,61 @@ var FieldAccuracyGrader = class {
     }
     if (withinTolerance) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: `${path50} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path51} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path51} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path50, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path51, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (unparseable candidate date)`
+        message: `${path51} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (unparseable expected date)`
+        message: `${path51} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: path50
+        message: path51
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path51} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -22670,11 +22676,11 @@ var FieldAccuracyGrader = class {
     };
   }
 };
-function resolvePath(obj, path50) {
-  if (!path50 || !obj) {
+function resolvePath(obj, path51) {
+  if (!path51 || !obj) {
     return void 0;
   }
-  const parts = path50.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path51.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -23205,8 +23211,8 @@ var TokenUsageGrader = class {
     };
   }
 };
-function getNestedValue(obj, path50) {
-  const parts = path50.split(".");
+function getNestedValue(obj, path51) {
+  const parts = path51.split(".");
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0 || typeof current !== "object") {
@@ -33200,6 +33206,47 @@ async function cleanupEvalWorkspaces(evalRunId, workspaceRoot) {
     await rm4(evalDir, { recursive: true, force: true });
   }
 }
+function interpolateArgs(args, context) {
+  const vars = {
+    workspace_path: context.workspacePath,
+    test_id: context.testId,
+    eval_run_id: context.evalRunId,
+    case_input: context.caseInput ?? "",
+    case_metadata: context.caseMetadata ? JSON.stringify(context.caseMetadata) : ""
+  };
+  return args.map((arg) => arg.replace(/\{\{(\w+)\}\}/g, (match, name) => vars[name] ?? match));
+}
+async function executeWorkspaceScript(config2, context, failureMode = "fatal") {
+  const stdin = JSON.stringify({
+    workspace_path: context.workspacePath,
+    test_id: context.testId,
+    eval_run_id: context.evalRunId,
+    case_input: context.caseInput ?? null,
+    case_metadata: context.caseMetadata ?? null
+  });
+  const timeoutMs = config2.timeout_ms ?? (failureMode === "fatal" ? 6e4 : 3e4);
+  const cwd = config2.cwd ?? context.workspaceFileDir ?? context.evalDir;
+  if (config2.script !== void 0 && config2.command === void 0) {
+    console.warn(
+      "\x1B[33mWarning: 'script' is deprecated in workspace config. Use 'command' instead.\x1B[0m"
+    );
+  }
+  const rawCommand = config2.command ?? config2.script ?? [];
+  const commandArray = interpolateArgs(rawCommand, context);
+  const result = await execFileWithStdin(commandArray, stdin, {
+    timeoutMs,
+    cwd
+  });
+  if (result.exitCode !== 0) {
+    const stderr = result.stderr.trim();
+    const message = stderr ? `${stderr}` : `Process exited with code ${result.exitCode}`;
+    if (failureMode === "fatal") {
+      throw new Error(`Script failed: ${message}`);
+    }
+    console.warn(`Script warning: ${message}`);
+  }
+  return result.stdout;
+}
 var GITHUB_SHORTHAND_RE = /^[A-Za-z0-9_.-]+\/[A-Za-z0-9_.-]+$/;
 function resolveRepoCloneUrl(repo) {
   const trimmed = repo.trim();
@@ -34184,46 +34231,919 @@ async function resolveWorkspaceTemplate(templatePath) {
   }
   return { dir: resolved };
 }
-function interpolateArgs(args, context) {
-  const vars = {
-    workspace_path: context.workspacePath,
-    test_id: context.testId,
-    eval_run_id: context.evalRunId,
-    case_input: context.caseInput ?? "",
-    case_metadata: context.caseMetadata ? JSON.stringify(context.caseMetadata) : ""
+var execFileAsync2 = promisify6(execFile2);
+var WORKSPACE_GIT_TIMEOUT_MS = 3e5;
+var WorkspaceSetupError = class extends Error {
+  failureStage;
+  failureReasonCode;
+  hookExecutions;
+  constructor(message, options) {
+    super(message);
+    this.name = "WorkspaceSetupError";
+    this.failureStage = options.failureStage;
+    this.failureReasonCode = options.failureReasonCode;
+    this.hookExecutions = options.hookExecutions ?? [];
+    if (options.cause !== void 0) {
+      this.cause = options.cause;
+    }
+  }
+};
+function toScriptConfig(hook, hookName, context) {
+  const command = hook.command ?? hook.script;
+  if (!command || command.length === 0) {
+    throw new Error(`${hookName} hook in ${context} requires command or script`);
+  }
+  return {
+    command,
+    ...hook.timeout_ms !== void 0 && { timeout_ms: hook.timeout_ms },
+    ...hook.timeoutMs !== void 0 && { timeoutMs: hook.timeoutMs },
+    ...hook.cwd !== void 0 && { cwd: hook.cwd },
+    ...hook.script !== void 0 && { script: hook.script }
   };
-  return args.map((arg) => arg.replace(/\{\{(\w+)\}\}/g, (match, name) => vars[name] ?? match));
 }
-async function executeWorkspaceScript(config2, context, failureMode = "fatal") {
-  const stdin = JSON.stringify({
-    workspace_path: context.workspacePath,
-    test_id: context.testId,
-    eval_run_id: context.evalRunId,
-    case_input: context.caseInput ?? null,
-    case_metadata: context.caseMetadata ?? null
-  });
-  const timeoutMs = config2.timeout_ms ?? (failureMode === "fatal" ? 6e4 : 3e4);
-  const cwd = config2.cwd ?? context.workspaceFileDir ?? context.evalDir;
-  if (config2.script !== void 0 && config2.command === void 0) {
+function hasHookCommand(hook) {
+  return !!(hook?.command && hook.command.length > 0 || hook?.script && hook.script.length > 0);
+}
+function hooksEnabled(workspace) {
+  return workspace?.hooks?.enabled !== false;
+}
+function workspaceGitEnv() {
+  const env = { ...process.env };
+  for (const key of Object.keys(env)) {
+    if (key.startsWith("GIT_") && key !== "GIT_SSH_COMMAND") {
+      delete env[key];
+    }
+  }
+  return {
+    ...env,
+    GIT_TERMINAL_PROMPT: "0",
+    GIT_ASKPASS: "",
+    GIT_SSH_COMMAND: "ssh -o BatchMode=yes"
+  };
+}
+async function resetWorkspaceRoot(workspacePath, resetMode, baselineRef) {
+  if (!existsSync6(path39.join(workspacePath, ".git"))) {
+    return false;
+  }
+  const cleanFlag = resetMode === "strict" ? "-fdx" : "-fd";
+  const opts = {
+    cwd: workspacePath,
+    timeout: WORKSPACE_GIT_TIMEOUT_MS,
+    env: workspaceGitEnv(),
+    maxBuffer: 50 * 1024 * 1024
+  };
+  await execFileAsync2("git", ["reset", "--hard", baselineRef ?? "HEAD"], opts);
+  await execFileAsync2("git", ["clean", cleanFlag], opts);
+  return true;
+}
+function commandForHook(hook) {
+  return hook?.command ?? hook?.script;
+}
+function hookExecution(options) {
+  const command = commandForHook(options.hook);
+  return {
+    scope: options.scope,
+    name: options.name,
+    status: options.status,
+    testId: options.testId,
+    ...options.workspacePath !== void 0 && { workspacePath: options.workspacePath },
+    ...command !== void 0 && { command },
+    ...options.hook?.cwd !== void 0 && { cwd: options.hook.cwd },
+    ...options.output !== void 0 && { output: options.output },
+    ...options.error !== void 0 && { error: options.error }
+  };
+}
+async function releasePoolSlots(setup) {
+  if (!setup.poolManager) {
+    return;
+  }
+  if (setup.poolSlot) {
+    await setup.poolManager.releaseSlot(setup.poolSlot);
+  }
+  for (const slot of setup.poolSlots) {
+    if (slot !== setup.poolSlot) {
+      await setup.poolManager.releaseSlot(slot).catch(() => {
+      });
+    }
+  }
+}
+async function releaseSharedWorkspaceSetup(setup) {
+  await releasePoolSlots(setup);
+}
+async function prepareSharedWorkspaceSetup(options) {
+  const {
+    evalRunId,
+    evalCases,
+    targetHooks,
+    evalDir,
+    verbose,
+    workers,
+    poolMaxSlots: configPoolMaxSlots,
+    workspacePath,
+    legacyWorkspacePath,
+    workspaceMode,
+    workspaceClean
+  } = options;
+  const suiteWorkspace = evalCases[0]?.workspace;
+  const rawTemplate = suiteWorkspace?.template;
+  const resolvedTemplate = await resolveWorkspaceTemplate(rawTemplate);
+  const workspaceTemplate = resolvedTemplate?.dir;
+  let suiteWorkspaceFile = resolvedTemplate?.workspaceFile;
+  const setupLog = (message) => {
+    if (verbose) {
+      console.log(`[setup] ${message}`);
+    }
+  };
+  const isPerTestIsolation = suiteWorkspace?.isolation === "per_test";
+  const cliWorkspacePath = workspacePath ?? legacyWorkspacePath;
+  const yamlWorkspacePath = suiteWorkspace?.path;
+  if (cliWorkspacePath && workspaceMode && workspaceMode !== "static") {
+    throw new Error("--workspace-path requires --workspace-mode static when both are provided");
+  }
+  let configuredMode = cliWorkspacePath ? "static" : workspaceMode ?? suiteWorkspace?.mode ?? (yamlWorkspacePath ? "static" : "pooled");
+  const configuredStaticPath = cliWorkspacePath ?? yamlWorkspacePath;
+  if (configuredMode === "static" && !configuredStaticPath) {
+    if (!suiteWorkspace?.repos?.length) {
+      setupLog("workspace.mode=static with no path and no repos \u2014 falling back to temp mode");
+      configuredMode = "temp";
+    } else {
+      throw new Error("workspace.mode=static requires workspace.path or --workspace-path");
+    }
+  }
+  const useStaticWorkspace = configuredMode === "static";
+  if (useStaticWorkspace && isPerTestIsolation) {
+    throw new Error(
+      "static workspace mode is incompatible with isolation: per_test. Use isolation: shared (default)."
+    );
+  }
+  if (configuredMode !== "static" && configuredStaticPath) {
+    throw new Error("workspace.path requires workspace.mode=static");
+  }
+  const hasSharedWorkspace = !!(useStaticWorkspace || !isPerTestIsolation && (workspaceTemplate || suiteWorkspace?.hooks || suiteWorkspace?.repos?.length));
+  const poolEnabled = configuredMode === "pooled";
+  const usePool = poolEnabled !== false && !!suiteWorkspace?.repos?.length && !isPerTestIsolation && !useStaticWorkspace;
+  setupLog(
+    `sharedWorkspace=${hasSharedWorkspace} perTestIsolation=${isPerTestIsolation} usePool=${usePool} workers=${workers}`
+  );
+  if (hasSharedWorkspace && !usePool && workers > 1 && evalCases.length > 1) {
     console.warn(
-      "\x1B[33mWarning: 'script' is deprecated in workspace config. Use 'command' instead.\x1B[0m"
+      [
+        `Warning: This eval uses a shared workspace with ${workers} workers.`,
+        "If the agent under test makes file edits, concurrent runs may corrupt each other.",
+        "To limit concurrency, add this to your eval YAML:",
+        "",
+        "  execution:",
+        "    workers: 1",
+        "",
+        "Or pass --workers 1 on the command line."
+      ].join("\n")
     );
   }
-  const rawCommand = config2.command ?? config2.script ?? [];
-  const commandArray = interpolateArgs(rawCommand, context);
-  const result = await execFileWithStdin(commandArray, stdin, {
-    timeoutMs,
-    cwd
-  });
-  if (result.exitCode !== 0) {
-    const stderr = result.stderr.trim();
-    const message = stderr ? `${stderr}` : `Process exited with code ${result.exitCode}`;
-    if (failureMode === "fatal") {
-      throw new Error(`Script failed: ${message}`);
+  let sharedWorkspacePath;
+  let sharedBaselineCommit;
+  let beforeAllOutput;
+  let poolManager;
+  let poolSlot;
+  const poolSlots = [];
+  const availablePoolSlots = [];
+  const poolSlotBaselines = /* @__PURE__ */ new Map();
+  const hookExecutions = [];
+  const poolMaxSlots = Math.min(configPoolMaxSlots ?? 10, 50);
+  let staticMaterialised = false;
+  const isYamlConfiguredPath = !cliWorkspacePath && !!yamlWorkspacePath;
+  let repoManager;
+  try {
+    if (useStaticWorkspace && configuredStaticPath) {
+      const dirExists = await stat8(configuredStaticPath).then(
+        (s) => s.isDirectory(),
+        () => false
+      );
+      const isEmpty = dirExists ? (await readdir8(configuredStaticPath)).length === 0 : false;
+      if (isYamlConfiguredPath && (!dirExists || isEmpty)) {
+        if (!dirExists) {
+          await mkdir17(configuredStaticPath, { recursive: true });
+        }
+        if (workspaceTemplate) {
+          await copyDirectoryRecursive(workspaceTemplate, configuredStaticPath);
+          setupLog(`copied template into static workspace: ${configuredStaticPath}`);
+        }
+        staticMaterialised = true;
+        setupLog(`materialised static workspace at: ${configuredStaticPath}`);
+      } else {
+        setupLog(`reusing existing static workspace: ${configuredStaticPath}`);
+      }
+      sharedWorkspacePath = configuredStaticPath;
+    } else if (!isPerTestIsolation && usePool && suiteWorkspace?.repos) {
+      const slotsNeeded = workers;
+      setupLog(`acquiring ${slotsNeeded} workspace pool slot(s) (pool capacity: ${poolMaxSlots})`);
+      poolManager = new WorkspacePoolManager(getWorkspacePoolRoot());
+      const poolRepoManager = new RepoManager(verbose);
+      for (let i = 0; i < slotsNeeded; i++) {
+        const slot = await poolManager.acquireWorkspace({
+          templatePath: workspaceTemplate,
+          repos: suiteWorkspace.repos,
+          maxSlots: poolMaxSlots,
+          repoManager: poolRepoManager,
+          poolReset: (workspaceClean === "full" ? "strict" : workspaceClean === "standard" ? "fast" : null) ?? "fast"
+        });
+        poolSlots.push(slot);
+        setupLog(`pool slot ${i} acquired at: ${slot.path} (existing=${slot.isExisting})`);
+      }
+      if (slotsNeeded === 1) {
+        poolSlot = poolSlots[0];
+        sharedWorkspacePath = poolSlot.path;
+      } else {
+        availablePoolSlots.push(...poolSlots);
+      }
+    } else if (!isPerTestIsolation && workspaceTemplate) {
+      setupLog(`creating shared workspace from template: ${workspaceTemplate}`);
+      try {
+        sharedWorkspacePath = await createTempWorkspace(workspaceTemplate, evalRunId, "shared");
+        setupLog(`shared workspace created at: ${sharedWorkspacePath}`);
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        throw new WorkspaceSetupError(`Failed to create shared workspace: ${message}`, {
+          failureStage: "setup",
+          failureReasonCode: "template_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    } else if (!isPerTestIsolation && (suiteWorkspace?.hooks || suiteWorkspace?.repos?.length)) {
+      sharedWorkspacePath = getWorkspacePath(evalRunId, "shared");
+      await mkdir17(sharedWorkspacePath, { recursive: true });
+      setupLog(`created empty shared workspace at: ${sharedWorkspacePath}`);
     }
-    console.warn(`Script warning: ${message}`);
+    if (suiteWorkspaceFile && sharedWorkspacePath) {
+      const copiedWorkspaceFile = path39.join(sharedWorkspacePath, path39.basename(suiteWorkspaceFile));
+      try {
+        await stat8(copiedWorkspaceFile);
+        suiteWorkspaceFile = copiedWorkspaceFile;
+      } catch {
+      }
+    }
+    const hasReposToMaterialize = !!suiteWorkspace?.repos?.length && !usePool && !isPerTestIsolation;
+    const needsRepoMaterialisation = hasReposToMaterialize && (!useStaticWorkspace || staticMaterialised);
+    const needsPerRepoCheck = hasReposToMaterialize && useStaticWorkspace && !staticMaterialised && isYamlConfiguredPath;
+    repoManager = needsRepoMaterialisation || needsPerRepoCheck ? new RepoManager(verbose) : void 0;
+    if (repoManager && sharedWorkspacePath && suiteWorkspace?.repos) {
+      try {
+        if (needsPerRepoCheck) {
+          for (const repo of suiteWorkspace.repos) {
+            if (!repo.path || !repo.repo) continue;
+            const targetDir = path39.join(sharedWorkspacePath, repo.path);
+            if (existsSync6(targetDir)) {
+              setupLog(`reusing existing repo at: ${targetDir}`);
+              continue;
+            }
+            setupLog(`materializing missing repo: ${repo.path}`);
+            await repoManager.materialize(repo, sharedWorkspacePath);
+          }
+        } else {
+          setupLog(
+            `materializing ${suiteWorkspace.repos.length} shared repo(s) into ${sharedWorkspacePath}`
+          );
+          await repoManager.materializeAll(suiteWorkspace.repos, sharedWorkspacePath);
+        }
+        setupLog("shared repo materialization complete");
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+          });
+        }
+        throw new WorkspaceSetupError(`Failed to materialize repos: ${message}`, {
+          failureStage: "repo_setup",
+          failureReasonCode: "clone_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+    const suiteDockerConfig = suiteWorkspace?.docker;
+    if (suiteDockerConfig) {
+      setupLog(`pulling Docker image: ${suiteDockerConfig.image}`);
+      const { DockerWorkspaceProvider } = await import("./docker-workspace-RPPXBT27-B4AQHVWA.js");
+      const dockerSetup = new DockerWorkspaceProvider(suiteDockerConfig);
+      if (!await dockerSetup.isDockerAvailable()) {
+        throw new Error(
+          "Docker workspace configured but Docker CLI is not available. Install Docker and ensure it is running."
+        );
+      }
+      await dockerSetup.pullImage();
+      setupLog("Docker image pull complete");
+    }
+    if (suiteWorkspace?.env) {
+      try {
+        await runPreflightChecks(suiteWorkspace.env, sharedWorkspacePath ?? void 0, setupLog);
+        setupLog("preflight checks passed");
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+          });
+        }
+        throw new WorkspaceSetupError(message, {
+          failureStage: "setup",
+          failureReasonCode: "preflight_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+    const suiteHooksEnabled = hooksEnabled(suiteWorkspace);
+    const suiteBeforeAllHook = suiteWorkspace?.hooks?.before_all;
+    if (sharedWorkspacePath && suiteHooksEnabled && hasHookCommand(suiteBeforeAllHook)) {
+      const beforeAllHook = suiteBeforeAllHook;
+      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
+      setupLog(
+        `running shared before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
+      );
+      const scriptContext = {
+        workspacePath: sharedWorkspacePath,
+        testId: "__before_all__",
+        evalRunId,
+        evalDir,
+        workspaceFileDir: suiteWorkspace?.workspaceFileDir
+      };
+      try {
+        beforeAllOutput = await executeWorkspaceScript(
+          toScriptConfig(beforeAllHook, "before_all", "suite workspace"),
+          scriptContext
+        );
+        hookExecutions.push(
+          hookExecution({
+            scope: "workspace",
+            name: "before_all",
+            status: "success",
+            testId: "__before_all__",
+            workspacePath: sharedWorkspacePath,
+            hook: beforeAllHook,
+            output: beforeAllOutput
+          })
+        );
+        setupLog("shared before_all completed");
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        hookExecutions.push(
+          hookExecution({
+            scope: "workspace",
+            name: "before_all",
+            status: "failed",
+            testId: "__before_all__",
+            workspacePath: sharedWorkspacePath,
+            hook: beforeAllHook,
+            error: message
+          })
+        );
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+          });
+        }
+        throw new WorkspaceSetupError(`before_all script failed: ${message}`, {
+          failureStage: "setup",
+          failureReasonCode: "script_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+    if (availablePoolSlots.length > 0 && suiteHooksEnabled && hasHookCommand(suiteBeforeAllHook)) {
+      const beforeAllHook = suiteBeforeAllHook;
+      for (const slot of availablePoolSlots) {
+        setupLog(`running before_all on pool slot ${slot.index}`);
+        const scriptContext = {
+          workspacePath: slot.path,
+          testId: "__before_all__",
+          evalRunId,
+          evalDir,
+          workspaceFileDir: suiteWorkspace?.workspaceFileDir
+        };
+        try {
+          const output = await executeWorkspaceScript(
+            toScriptConfig(beforeAllHook, "before_all", "suite workspace"),
+            scriptContext
+          );
+          if (!beforeAllOutput) beforeAllOutput = output;
+          hookExecutions.push(
+            hookExecution({
+              scope: "workspace",
+              name: "before_all",
+              status: "success",
+              testId: "__before_all__",
+              workspacePath: slot.path,
+              hook: beforeAllHook,
+              output
+            })
+          );
+          setupLog(`before_all completed on pool slot ${slot.index}`);
+        } catch (error40) {
+          const message = error40 instanceof Error ? error40.message : String(error40);
+          hookExecutions.push(
+            hookExecution({
+              scope: "workspace",
+              name: "before_all",
+              status: "failed",
+              testId: "__before_all__",
+              workspacePath: slot.path,
+              hook: beforeAllHook,
+              error: message
+            })
+          );
+          throw new WorkspaceSetupError(
+            `before_all script failed on pool slot ${slot.index}: ${message}`,
+            {
+              failureStage: "setup",
+              failureReasonCode: "script_error",
+              hookExecutions,
+              cause: error40
+            }
+          );
+        }
+      }
+    }
+    const targetBeforeAllHook = targetHooks?.before_all;
+    if (sharedWorkspacePath && hasHookCommand(targetBeforeAllHook)) {
+      const beforeAllCommand = (targetBeforeAllHook.command ?? []).join(" ");
+      setupLog(`running target before_all command=${beforeAllCommand}`);
+      const scriptContext = {
+        workspacePath: sharedWorkspacePath,
+        testId: "__target_before_all__",
+        evalRunId,
+        evalDir,
+        workspaceFileDir: suiteWorkspace?.workspaceFileDir
+      };
+      try {
+        await executeWorkspaceScript(
+          toScriptConfig(targetBeforeAllHook, "before_all", "target hooks"),
+          scriptContext
+        );
+        hookExecutions.push(
+          hookExecution({
+            scope: "target",
+            name: "before_all",
+            status: "success",
+            testId: "__target_before_all__",
+            workspacePath: sharedWorkspacePath,
+            hook: targetBeforeAllHook
+          })
+        );
+        setupLog("target before_all completed");
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        hookExecutions.push(
+          hookExecution({
+            scope: "target",
+            name: "before_all",
+            status: "failed",
+            testId: "__target_before_all__",
+            workspacePath: sharedWorkspacePath,
+            hook: targetBeforeAllHook,
+            error: message
+          })
+        );
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+          });
+        }
+        throw new WorkspaceSetupError(`target before_all hook failed: ${message}`, {
+          failureStage: "setup",
+          failureReasonCode: "script_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+    if (availablePoolSlots.length > 0 && hasHookCommand(targetBeforeAllHook)) {
+      for (const slot of availablePoolSlots) {
+        setupLog(`running target before_all on pool slot ${slot.index}`);
+        const scriptContext = {
+          workspacePath: slot.path,
+          testId: "__target_before_all__",
+          evalRunId,
+          evalDir,
+          workspaceFileDir: suiteWorkspace?.workspaceFileDir
+        };
+        try {
+          await executeWorkspaceScript(
+            toScriptConfig(targetBeforeAllHook, "before_all", "target hooks"),
+            scriptContext
+          );
+          hookExecutions.push(
+            hookExecution({
+              scope: "target",
+              name: "before_all",
+              status: "success",
+              testId: "__target_before_all__",
+              workspacePath: slot.path,
+              hook: targetBeforeAllHook
+            })
+          );
+        } catch (error40) {
+          const message = error40 instanceof Error ? error40.message : String(error40);
+          hookExecutions.push(
+            hookExecution({
+              scope: "target",
+              name: "before_all",
+              status: "failed",
+              testId: "__target_before_all__",
+              workspacePath: slot.path,
+              hook: targetBeforeAllHook,
+              error: message
+            })
+          );
+          throw new WorkspaceSetupError(
+            `target before_all hook failed on pool slot ${slot.index}: ${message}`,
+            {
+              failureStage: "setup",
+              failureReasonCode: "script_error",
+              hookExecutions,
+              cause: error40
+            }
+          );
+        }
+      }
+    }
+    if (sharedWorkspacePath) {
+      try {
+        sharedBaselineCommit = await initializeBaseline(sharedWorkspacePath);
+        setupLog(`shared baseline initialized: ${sharedBaselineCommit}`);
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        setupLog(`shared baseline initialization failed (file_changes unavailable): ${message}`);
+      }
+    }
+    if (availablePoolSlots.length > 0) {
+      for (const slot of availablePoolSlots) {
+        try {
+          const baseline = await initializeBaseline(slot.path);
+          poolSlotBaselines.set(slot.path, baseline);
+          setupLog(`pool slot ${slot.index} baseline initialized: ${baseline}`);
+        } catch (error40) {
+          const message = error40 instanceof Error ? error40.message : String(error40);
+          setupLog(
+            `pool slot ${slot.index} baseline initialization failed (file_changes unavailable): ${message}`
+          );
+        }
+      }
+    }
+    return {
+      ...suiteWorkspace !== void 0 && { suiteWorkspace },
+      ...sharedWorkspacePath !== void 0 && { sharedWorkspacePath },
+      ...sharedBaselineCommit !== void 0 && { sharedBaselineCommit },
+      ...suiteWorkspaceFile !== void 0 && { suiteWorkspaceFile },
+      ...beforeAllOutput !== void 0 && { beforeAllOutput },
+      ...repoManager !== void 0 && { repoManager },
+      ...poolManager !== void 0 && { poolManager },
+      ...poolSlot !== void 0 && { poolSlot },
+      poolSlots,
+      availablePoolSlots,
+      poolSlotBaselines,
+      useStaticWorkspace,
+      configuredMode,
+      hookExecutions
+    };
+  } catch (error40) {
+    await releasePoolSlots({ poolManager, poolSlot, poolSlots }).catch(() => {
+    });
+    throw error40;
+  }
+}
+async function prepareEvalCaseWorkspace(options) {
+  const {
+    evalCase,
+    evalRunId,
+    sharedWorkspacePath,
+    sharedBaselineCommit,
+    suiteWorkspaceFile,
+    repoManager,
+    evalDir,
+    cleanupWorkspaces: forceCleanup,
+    targetHooks,
+    setupDebug
+  } = options;
+  let workspacePath = sharedWorkspacePath;
+  let beforeAllOutput;
+  let beforeEachOutput;
+  const isSharedWorkspace = !!sharedWorkspacePath;
+  let caseWorkspaceFile;
+  const caseHooksEnabled = hooksEnabled(evalCase.workspace);
+  const hookExecutions = [];
+  if (!workspacePath) {
+    const rawCaseTemplate = evalCase.workspace?.template;
+    const resolvedCaseTemplate = await resolveWorkspaceTemplate(rawCaseTemplate);
+    const caseWorkspaceTemplate = resolvedCaseTemplate?.dir;
+    caseWorkspaceFile = resolvedCaseTemplate?.workspaceFile;
+    if (caseWorkspaceTemplate && evalRunId) {
+      try {
+        workspacePath = await createTempWorkspace(caseWorkspaceTemplate, evalRunId, evalCase.id);
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        throw new WorkspaceSetupError(`Failed to create workspace: ${message}`, {
+          failureStage: "setup",
+          failureReasonCode: "template_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+      if (caseWorkspaceFile && workspacePath) {
+        const copiedFile = path39.join(workspacePath, path39.basename(caseWorkspaceFile));
+        try {
+          await stat8(copiedFile);
+          caseWorkspaceFile = copiedFile;
+        } catch {
+        }
+      }
+    }
+    if (!workspacePath && (evalCase.workspace?.hooks || evalCase.workspace?.repos?.length) && evalRunId) {
+      workspacePath = getWorkspacePath(evalRunId, evalCase.id);
+      await mkdir17(workspacePath, { recursive: true });
+    }
+    if (evalCase.workspace?.repos?.length && workspacePath) {
+      const perCaseRepoManager = new RepoManager(setupDebug);
+      try {
+        if (setupDebug) {
+          console.log(
+            `[setup] test=${evalCase.id} materializing ${evalCase.workspace.repos.length} per-test repo(s) into ${workspacePath}`
+          );
+        }
+        await perCaseRepoManager.materializeAll(evalCase.workspace.repos, workspacePath);
+        if (setupDebug) {
+          console.log(`[setup] test=${evalCase.id} per-test repo materialization complete`);
+        }
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        throw new WorkspaceSetupError(`Failed to materialize repos: ${message}`, {
+          failureStage: "repo_setup",
+          failureReasonCode: "clone_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+    if (workspacePath && evalCase.metadata?.agent_skills_files) {
+      const baseDir = evalCase.metadata.agent_skills_base_dir;
+      const files = evalCase.metadata.agent_skills_files;
+      if (baseDir && files.length > 0) {
+        for (const relPath of files) {
+          const srcPath = path39.resolve(baseDir, relPath);
+          const destPath = path39.resolve(workspacePath, relPath);
+          try {
+            await mkdir17(path39.dirname(destPath), { recursive: true });
+            await copyFile2(srcPath, destPath);
+          } catch (error40) {
+            const message = error40 instanceof Error ? error40.message : String(error40);
+            throw new WorkspaceSetupError(
+              `Agent Skills eval file not found: ${relPath} (resolved from ${baseDir}): ${message}`,
+              {
+                failureStage: "setup",
+                failureReasonCode: "file_copy_error",
+                hookExecutions,
+                cause: error40
+              }
+            );
+          }
+        }
+      }
+    }
+    const caseBeforeAllHook = evalCase.workspace?.hooks?.before_all;
+    if (workspacePath && caseHooksEnabled && hasHookCommand(caseBeforeAllHook)) {
+      const beforeAllHook = caseBeforeAllHook;
+      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
+      if (setupDebug) {
+        console.log(
+          `[setup] test=${evalCase.id} running before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
+        );
+      }
+      const scriptContext = {
+        workspacePath,
+        testId: evalCase.id,
+        evalRunId: evalRunId ?? "",
+        caseInput: evalCase.question,
+        caseMetadata: evalCase.metadata,
+        evalDir,
+        workspaceFileDir: evalCase.workspace?.workspaceFileDir
+      };
+      try {
+        beforeAllOutput = await executeWorkspaceScript(
+          toScriptConfig(beforeAllHook, "before_all", `test '${evalCase.id}'`),
+          scriptContext
+        );
+        hookExecutions.push(
+          hookExecution({
+            scope: "workspace",
+            name: "before_all",
+            status: "success",
+            testId: evalCase.id,
+            workspacePath,
+            hook: beforeAllHook,
+            output: beforeAllOutput
+          })
+        );
+        if (setupDebug) {
+          console.log(`[setup] test=${evalCase.id} before_all completed`);
+        }
+      } catch (error40) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        hookExecutions.push(
+          hookExecution({
+            scope: "workspace",
+            name: "before_all",
+            status: "failed",
+            testId: evalCase.id,
+            workspacePath,
+            hook: beforeAllHook,
+            error: message
+          })
+        );
+        if (forceCleanup && workspacePath) {
+          await cleanupWorkspace(workspacePath).catch(() => {
+          });
+        }
+        throw new WorkspaceSetupError(`before_all script failed: ${message}`, {
+          failureStage: "setup",
+          failureReasonCode: "script_error",
+          hookExecutions,
+          cause: error40
+        });
+      }
+    }
+  }
+  let beforeEachNeedsFreshBaseline = false;
+  if (caseHooksEnabled && workspacePath && evalCase.workspace?.hooks?.before_each?.reset && evalCase.workspace.hooks.before_each.reset !== "none") {
+    try {
+      if (repoManager && evalCase.workspace.repos?.length) {
+        await repoManager.reset(
+          evalCase.workspace.repos,
+          workspacePath,
+          evalCase.workspace.hooks.before_each.reset
+        );
+      } else {
+        await resetWorkspaceRoot(
+          workspacePath,
+          evalCase.workspace.hooks.before_each.reset,
+          sharedBaselineCommit
+        );
+      }
+    } catch (error40) {
+      const message = error40 instanceof Error ? error40.message : String(error40);
+      throw new WorkspaceSetupError(`before_each reset failed: ${message}`, {
+        failureStage: "setup",
+        failureReasonCode: "script_error",
+        hookExecutions,
+        cause: error40
+      });
+    }
+  }
+  const caseBeforeEachHook = evalCase.workspace?.hooks?.before_each;
+  if (workspacePath && caseHooksEnabled && hasHookCommand(caseBeforeEachHook)) {
+    const beforeEachHook = caseBeforeEachHook;
+    const scriptContext = {
+      workspacePath,
+      testId: evalCase.id,
+      evalRunId: evalRunId ?? "",
+      caseInput: evalCase.question,
+      caseMetadata: evalCase.metadata,
+      evalDir,
+      workspaceFileDir: evalCase.workspace?.workspaceFileDir
+    };
+    try {
+      beforeEachOutput = await executeWorkspaceScript(
+        toScriptConfig(beforeEachHook, "before_each", `test '${evalCase.id}'`),
+        scriptContext
+      );
+      hookExecutions.push(
+        hookExecution({
+          scope: "workspace",
+          name: "before_each",
+          status: "success",
+          testId: evalCase.id,
+          workspacePath,
+          hook: beforeEachHook,
+          output: beforeEachOutput
+        })
+      );
+      beforeEachNeedsFreshBaseline = true;
+    } catch (error40) {
+      const message = error40 instanceof Error ? error40.message : String(error40);
+      hookExecutions.push(
+        hookExecution({
+          scope: "workspace",
+          name: "before_each",
+          status: "failed",
+          testId: evalCase.id,
+          workspacePath,
+          hook: beforeEachHook,
+          error: message
+        })
+      );
+      throw new WorkspaceSetupError(`before_each script failed: ${message}`, {
+        failureStage: "setup",
+        failureReasonCode: "script_error",
+        hookExecutions,
+        cause: error40
+      });
+    }
+  }
+  const targetBeforeEachHook = targetHooks?.before_each;
+  if (workspacePath && hasHookCommand(targetBeforeEachHook)) {
+    const scriptContext = {
+      workspacePath,
+      testId: evalCase.id,
+      evalRunId: evalRunId ?? "",
+      caseInput: evalCase.question,
+      caseMetadata: evalCase.metadata,
+      evalDir,
+      workspaceFileDir: evalCase.workspace?.workspaceFileDir
+    };
+    try {
+      await executeWorkspaceScript(
+        toScriptConfig(targetBeforeEachHook, "before_each", `target hook for '${evalCase.id}'`),
+        scriptContext
+      );
+      hookExecutions.push(
+        hookExecution({
+          scope: "target",
+          name: "before_each",
+          status: "success",
+          testId: evalCase.id,
+          workspacePath,
+          hook: targetBeforeEachHook
+        })
+      );
+      beforeEachNeedsFreshBaseline = true;
+    } catch (error40) {
+      const message = error40 instanceof Error ? error40.message : String(error40);
+      hookExecutions.push(
+        hookExecution({
+          scope: "target",
+          name: "before_each",
+          status: "failed",
+          testId: evalCase.id,
+          workspacePath,
+          hook: targetBeforeEachHook,
+          error: message
+        })
+      );
+      throw new WorkspaceSetupError(`target before_each hook failed: ${message}`, {
+        failureStage: "setup",
+        failureReasonCode: "script_error",
+        hookExecutions,
+        cause: error40
+      });
+    }
+  }
+  let baselineCommit = beforeEachNeedsFreshBaseline ? void 0 : sharedBaselineCommit;
+  if (!baselineCommit && workspacePath) {
+    try {
+      baselineCommit = await initializeBaseline(workspacePath);
+    } catch (error40) {
+      const message = error40 instanceof Error ? error40.message : String(error40);
+      if (setupDebug) {
+        console.warn(`[setup] test=${evalCase.id} baseline initialization failed: ${message}`);
+      }
+    }
+  }
+  return {
+    ...workspacePath !== void 0 && { workspacePath },
+    caseWorkspaceFile: caseWorkspaceFile ?? suiteWorkspaceFile,
+    ...beforeAllOutput !== void 0 && { beforeAllOutput },
+    ...beforeEachOutput !== void 0 && { beforeEachOutput },
+    ...baselineCommit !== void 0 && { baselineCommit },
+    isSharedWorkspace,
+    hookExecutions
+  };
+}
+async function runPreflightChecks(env, cwd, log) {
+  const missing = [];
+  for (const cmd of env.required_commands ?? []) {
+    log(`preflight: checking command "${cmd}"`);
+    try {
+      if (process.platform === "win32") {
+        await execFileAsync2("where", [cmd], { cwd });
+      } else {
+        await execFileAsync2("sh", ["-c", `command -v ${cmd}`], { cwd });
+      }
+    } catch {
+      missing.push(`command: ${cmd}`);
+    }
+  }
+  for (const mod of env.required_python_modules ?? []) {
+    log(`preflight: checking Python module "${mod}"`);
+    try {
+      await execFileAsync2("python3", ["-c", `import ${mod}`], { cwd });
+    } catch {
+      missing.push(`python module: ${mod}`);
+    }
+  }
+  if (missing.length > 0) {
+    throw new Error(
+      `Preflight checks failed \u2014 missing dependencies:
+${missing.map((m) => `  \u2022 ${m}`).join("\n")}
+
+Install the missing dependencies before running this eval.`
+    );
   }
-  return result.stdout;
 }
 function flattenInputMessages(messages) {
   return messages.flatMap((message) => extractContentSegments(message.content));
@@ -34305,7 +35225,7 @@ async function loadTestsFromAgentSkills(filePath) {
   } catch {
     throw new Error(`Invalid Agent Skills evals.json: failed to parse JSON in '${filePath}'`);
   }
-  return parseAgentSkillsEvals(parsed, filePath, path39.dirname(path39.resolve(filePath)));
+  return parseAgentSkillsEvals(parsed, filePath, path40.dirname(path40.resolve(filePath)));
 }
 function parseAgentSkillsEvals(parsed, source = "evals.json", baseDir) {
   if (!isAgentSkillsFormat(parsed)) {
@@ -34343,7 +35263,7 @@ function parseAgentSkillsEvals(parsed, source = "evals.json", baseDir) {
       if (baseDir) {
         metadata.agent_skills_base_dir = baseDir;
         for (const file2 of evalCase.files) {
-          filePaths.push(path39.resolve(baseDir, file2));
+          filePaths.push(path40.resolve(baseDir, file2));
         }
       }
     }
@@ -34379,15 +35299,15 @@ function resolveToAbsolutePath(candidate) {
     if (candidate.startsWith("file:")) {
       return fileURLToPath4(candidate);
     }
-    return path40.resolve(candidate);
+    return path41.resolve(candidate);
   }
   throw new TypeError("Unsupported repoRoot value. Expected string or URL.");
 }
 function buildDirectoryChain2(filePath, repoRoot) {
   const directories = [];
   const seen = /* @__PURE__ */ new Set();
-  const boundary = path40.resolve(repoRoot);
-  let current = path40.resolve(path40.dirname(filePath));
+  const boundary = path41.resolve(repoRoot);
+  let current = path41.resolve(path41.dirname(filePath));
   while (current !== void 0) {
     if (!seen.has(current)) {
       directories.push(current);
@@ -34396,7 +35316,7 @@ function buildDirectoryChain2(filePath, repoRoot) {
     if (current === boundary) {
       break;
     }
-    const parent = path40.dirname(current);
+    const parent = path41.dirname(current);
     if (parent === current) {
       break;
     }
@@ -34410,16 +35330,16 @@ function buildDirectoryChain2(filePath, repoRoot) {
 function buildSearchRoots2(evalPath, repoRoot) {
   const uniqueRoots = [];
   const addRoot = (root) => {
-    const normalized = path40.resolve(root);
+    const normalized = path41.resolve(root);
     if (!uniqueRoots.includes(normalized)) {
       uniqueRoots.push(normalized);
     }
   };
-  let currentDir = path40.dirname(evalPath);
+  let currentDir = path41.dirname(evalPath);
   let reachedBoundary = false;
   while (!reachedBoundary) {
     addRoot(currentDir);
-    const parentDir = path40.dirname(currentDir);
+    const parentDir = path41.dirname(currentDir);
     if (currentDir === repoRoot || parentDir === currentDir) {
       reachedBoundary = true;
     } else {
@@ -34437,16 +35357,16 @@ function trimLeadingSeparators2(value) {
 async function resolveFileReference3(rawValue, searchRoots) {
   const displayPath = trimLeadingSeparators2(rawValue);
   const potentialPaths = [];
-  if (path40.isAbsolute(rawValue)) {
-    potentialPaths.push(path40.normalize(rawValue));
+  if (path41.isAbsolute(rawValue)) {
+    potentialPaths.push(path41.normalize(rawValue));
   }
   for (const base of searchRoots) {
-    potentialPaths.push(path40.resolve(base, displayPath));
+    potentialPaths.push(path41.resolve(base, displayPath));
   }
   const attempted = [];
   const seen = /* @__PURE__ */ new Set();
   for (const candidate of potentialPaths) {
-    const absoluteCandidate = path40.resolve(candidate);
+    const absoluteCandidate = path41.resolve(candidate);
     if (seen.has(absoluteCandidate)) {
       continue;
     }
@@ -34467,9 +35387,9 @@ var DEFAULT_EVAL_PATTERNS = [
 ];
 async function loadConfig(evalFilePath, repoRoot) {
   const directories = buildDirectoryChain2(evalFilePath, repoRoot);
-  const globalConfigPath = path41.join(getAgentvConfigDir(), "config.yaml");
+  const globalConfigPath = path422.join(getAgentvConfigDir(), "config.yaml");
   for (const directory of directories) {
-    const configPath2 = path41.join(directory, ".agentv", "config.yaml");
+    const configPath2 = path422.join(directory, ".agentv", "config.yaml");
     if (!await fileExists3(configPath2)) {
       continue;
     }
@@ -35020,8 +35940,8 @@ function isTemplateReference(value) {
 }
 async function resolveAssertionTemplateReference(include, searchRoots) {
   const templateCandidates = isTemplateReference(include) ? [
-    path422.join(".agentv", "templates", `${include}.yaml`),
-    path422.join(".agentv", "templates", `${include}.yml`)
+    path43.join(".agentv", "templates", `${include}.yaml`),
+    path43.join(".agentv", "templates", `${include}.yml`)
   ] : [include];
   const attempted = [];
   for (const candidate of templateCandidates) {
@@ -35074,10 +35994,10 @@ ${resolved.attempted.map((attempt) => `  Tried: ${attempt}`).join("\n")}` : "";
       `Invalid assertion template file in '${evalId}': ${resolved.resolvedPath} is missing a top-level assertions array`
     );
   }
-  const templateDir = path422.dirname(resolved.resolvedPath);
+  const templateDir = path43.dirname(resolved.resolvedPath);
   const nestedSearchRoots = [
     templateDir,
-    ...searchRoots.filter((root) => path422.resolve(root) !== templateDir)
+    ...searchRoots.filter((root) => path43.resolve(root) !== templateDir)
   ];
   return await expandGraderEntries(assertions, nestedSearchRoots, evalId, {
     depth: nextDepth,
@@ -35138,7 +36058,7 @@ async function collectAssertionTemplateReferencesFromValue(value, searchRoots, e
         references.push({
           kind: "assertion_template",
           displayPath: resolved.displayPath,
-          ...resolved.resolvedPath ? { resolvedPath: path422.resolve(resolved.resolvedPath) } : {}
+          ...resolved.resolvedPath ? { resolvedPath: path43.resolve(resolved.resolvedPath) } : {}
         });
         if (resolved.resolvedPath) {
           if (includeContext.chain.includes(resolved.resolvedPath)) {
@@ -35148,10 +36068,10 @@ async function collectAssertionTemplateReferencesFromValue(value, searchRoots, e
           const content = await readFile15(resolved.resolvedPath, "utf8");
           const parsed = interpolateEnv(parseYamlValue(content), process.env);
           if (isJsonObject2(parsed) && Array.isArray(parsed.assertions)) {
-            const templateDir = path422.dirname(resolved.resolvedPath);
+            const templateDir = path43.dirname(resolved.resolvedPath);
             const nestedSearchRoots = [
               templateDir,
-              ...searchRoots.filter((root) => path422.resolve(root) !== templateDir)
+              ...searchRoots.filter((root) => path43.resolve(root) !== templateDir)
             ];
             references.push(
               ...await collectAssertionTemplateReferencesFromValue(
@@ -35337,7 +36257,7 @@ async function parseGraderList(candidateEvaluators, searchRoots, evalId, default
       if (cwd) {
         const resolved = await resolveFileReference3(cwd, searchRoots);
         if (resolved.resolvedPath) {
-          resolvedCwd = path422.resolve(resolved.resolvedPath);
+          resolvedCwd = path43.resolve(resolved.resolvedPath);
         } else {
           logWarning2(
             `Code-grader evaluator '${name}' in '${evalId}': cwd not found (${resolved.displayPath})`,
@@ -35523,7 +36443,7 @@ async function parseGraderList(candidateEvaluators, searchRoots, evalId, default
             aggregatorPrompt = fileRef;
             const resolved = await resolveFileReference3(fileRef, searchRoots);
             if (resolved.resolvedPath) {
-              promptPath2 = path422.resolve(resolved.resolvedPath);
+              promptPath2 = path43.resolve(resolved.resolvedPath);
             } else {
               throw new Error(
                 `Composite aggregator in '${evalId}': prompt file not found: ${resolved.displayPath}`
@@ -36203,7 +37123,7 @@ async function parseGraderList(candidateEvaluators, searchRoots, evalId, default
       const commandPath = commandArray[commandArray.length - 1];
       const resolved = await resolveFileReference3(commandPath, searchRoots);
       if (resolved.resolvedPath) {
-        resolvedPromptScript = [...commandArray.slice(0, -1), path422.resolve(resolved.resolvedPath)];
+        resolvedPromptScript = [...commandArray.slice(0, -1), path43.resolve(resolved.resolvedPath)];
       } else {
         throw new Error(
           `Grader '${name}' in '${evalId}': prompt command file not found: ${resolved.displayPath}`
@@ -36218,7 +37138,7 @@ async function parseGraderList(candidateEvaluators, searchRoots, evalId, default
         prompt = fileRef;
         const resolved = await resolveFileReference3(fileRef, searchRoots);
         if (resolved.resolvedPath) {
-          promptPath = path422.resolve(resolved.resolvedPath);
+          promptPath = path43.resolve(resolved.resolvedPath);
           try {
             await validateCustomPromptContent(promptPath);
           } catch (error40) {
@@ -36376,7 +37296,7 @@ async function parsePreprocessors(rawValue, searchRoots, evaluatorName, evalId)
     preprocessors.push({
       type,
       command,
-      resolvedCommand: [...command.slice(0, -1), path422.resolve(resolved.resolvedPath)]
+      resolvedCommand: [...command.slice(0, -1), path43.resolve(resolved.resolvedPath)]
     });
   }
   return preprocessors;
@@ -36471,10 +37391,10 @@ async function resolveOptionalCommandSource(command, searchRoots) {
     return void 0;
   }
   const resolved = await resolveFileReference3(candidate, searchRoots);
-  return resolved.resolvedPath ? path422.resolve(resolved.resolvedPath) : void 0;
+  return resolved.resolvedPath ? path43.resolve(resolved.resolvedPath) : void 0;
 }
 function looksLikeFilePath(value) {
-  return path422.isAbsolute(value) || value.startsWith(".") || value.includes("/") || value.includes("\\") || /\.[cm]?[jt]sx?$|\.py$|\.sh$|\.bash$|\.rb$|\.go$|\.rs$/i.test(value);
+  return path43.isAbsolute(value) || value.startsWith(".") || value.includes("/") || value.includes("\\") || /\.[cm]?[jt]sx?$|\.py$|\.sh$|\.bash$|\.rb$|\.go$|\.rs$/i.test(value);
 }
 function parseCommandToArgv(command) {
   if (process.platform === "win32") {
@@ -36862,7 +37782,7 @@ var IMAGE_MEDIA_TYPES = {
   ".bmp": "image/bmp"
 };
 function detectImageMediaType(filePath) {
-  const ext = path43.extname(filePath).toLowerCase();
+  const ext = path44.extname(filePath).toLowerCase();
   return IMAGE_MEDIA_TYPES[ext];
 }
 var ANSI_YELLOW3 = "\x1B[33m";
@@ -36926,7 +37846,7 @@ async function processMessages(options) {
             ...cloneJsonObject(rawSegment),
             path: displayPath,
             text: fileContent,
-            resolvedPath: path43.resolve(resolvedPath)
+            resolvedPath: path44.resolve(resolvedPath)
           });
           if (verbose) {
             const label = messageType === "input" ? "[File]" : "[Expected Output File]";
@@ -37050,7 +37970,7 @@ async function processExpectedMessages(options) {
               type: "file",
               path: displayPath,
               text: fileContent,
-              resolvedPath: path43.resolve(resolvedPath)
+              resolvedPath: path44.resolve(resolvedPath)
             });
             if (verbose) {
               console.log(`  [Expected Output File] Found: ${displayPath}`);
@@ -37196,7 +38116,7 @@ function matchesFilter(id, filter) {
   return typeof filter === "string" ? micromatch.isMatch(id, filter) : filter.some((pattern) => micromatch.isMatch(id, pattern));
 }
 function detectFormat(filePath) {
-  const ext = path44.extname(filePath).toLowerCase();
+  const ext = path45.extname(filePath).toLowerCase();
   if (ext === ".jsonl") return "jsonl";
   if (ext === ".yaml" || ext === ".yml") return "yaml";
   if (ext === ".json") return "agent-skills-json";
@@ -37206,9 +38126,9 @@ function detectFormat(filePath) {
   );
 }
 async function loadSidecarMetadata(jsonlPath, verbose) {
-  const dir = path44.dirname(jsonlPath);
-  const base = path44.basename(jsonlPath, ".jsonl");
-  const sidecarPath = path44.join(dir, `${base}.yaml`);
+  const dir = path45.dirname(jsonlPath);
+  const base = path45.basename(jsonlPath, ".jsonl");
+  const sidecarPath = path45.join(dir, `${base}.yaml`);
   if (!await fileExists3(sidecarPath)) {
     if (verbose) {
       logWarning4(`Sidecar metadata file not found: ${sidecarPath} (using defaults)`);
@@ -37257,13 +38177,13 @@ function parseJsonlContent(content, filePath) {
 async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
   const verbose = options?.verbose ?? false;
   const filterPattern = options?.filter;
-  const absoluteTestPath = path44.resolve(evalFilePath);
+  const absoluteTestPath = path45.resolve(evalFilePath);
   const repoRootPath = resolveToAbsolutePath(repoRoot);
   const searchRoots = buildSearchRoots2(absoluteTestPath, repoRootPath);
   const sidecar = await loadSidecarMetadata(absoluteTestPath, verbose);
   const rawFile = await readFile17(absoluteTestPath, "utf8");
   const rawCases = parseJsonlContent(rawFile, evalFilePath);
-  const fallbackSuiteName = path44.basename(absoluteTestPath, ".jsonl") || "eval";
+  const fallbackSuiteName = path45.basename(absoluteTestPath, ".jsonl") || "eval";
   const suiteName = sidecar.name && sidecar.name.trim().length > 0 ? sidecar.name : fallbackSuiteName;
   const globalEvaluator = coerceEvaluator(sidecar.evaluator, "sidecar") ?? "llm-grader";
   const globalExecution = sidecar.execution;
@@ -37660,7 +38580,7 @@ function interpolateRawEvalCase(raw, vars) {
 }
 async function readTestSuiteMetadata(testFilePath) {
   try {
-    const absolutePath = path45.resolve(testFilePath);
+    const absolutePath = path46.resolve(testFilePath);
     const content = await readFile18(absolutePath, "utf8");
     const parsed = interpolateEnv(parseYamlValue(content), process.env);
     if (!isJsonObject(parsed)) {
@@ -37685,7 +38605,7 @@ async function loadTestSuite(evalFilePath, repoRoot, options) {
     return { tests: await loadTestsFromAgentSkills(evalFilePath) };
   }
   if (format === "typescript") {
-    const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-NWH3B4HG-UXXCZKLP.js");
+    const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-ZVL6CGTE-TZYZX3QS.js");
     return loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
   }
   const { tests, parsed, suiteWorkspacePath } = await loadTestsFromYaml(
@@ -37720,7 +38640,7 @@ async function loadTests(evalFilePath, repoRoot, options) {
     return loadTestsFromAgentSkills(evalFilePath);
   }
   if (format === "typescript") {
-    const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-NWH3B4HG-UXXCZKLP.js");
+    const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-ZVL6CGTE-TZYZX3QS.js");
     const suite = await loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
     return suite.tests;
   }
@@ -37731,7 +38651,7 @@ var loadEvalCases = loadTests;
 async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
   const verbose = options?.verbose ?? false;
   const filterPattern = options?.filter;
-  const absoluteTestPath = path45.resolve(evalFilePath);
+  const absoluteTestPath = path46.resolve(evalFilePath);
   const repoRootPath = resolveToAbsolutePath(repoRoot);
   const searchRoots = buildSearchRoots2(absoluteTestPath, repoRootPath);
   const config2 = await loadConfig(absoluteTestPath, repoRootPath);
@@ -37744,7 +38664,7 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
   }
   const suite = interpolated;
   const suiteNameFromFile = asString5(suite.name)?.trim();
-  const fallbackSuiteName = path45.basename(absoluteTestPath).replace(/\.eval\.ya?ml$/i, "").replace(/\.ya?ml$/i, "") || "eval";
+  const fallbackSuiteName = path46.basename(absoluteTestPath).replace(/\.eval\.ya?ml$/i, "").replace(/\.ya?ml$/i, "") || "eval";
   const suiteName = suiteNameFromFile && suiteNameFromFile.length > 0 ? suiteNameFromFile : fallbackSuiteName;
   const rawTestCases = resolveTests(suite);
   const globalEvaluator = coerceEvaluator(suite.evaluator, "global") ?? "llm-grader";
@@ -37754,13 +38674,13 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
     "<suite>",
     absoluteTestPath
   );
-  const evalFileDir = path45.dirname(absoluteTestPath);
+  const evalFileDir = path46.dirname(absoluteTestPath);
   let expandedTestCases;
   if (typeof rawTestCases === "string") {
-    const externalPath = path45.resolve(evalFileDir, rawTestCases);
+    const externalPath = path46.resolve(evalFileDir, rawTestCases);
     let isDir = false;
     try {
-      const pathStat = await stat8(externalPath);
+      const pathStat = await stat9(externalPath);
       isDir = pathStat.isDirectory();
     } catch {
     }
@@ -38061,7 +38981,7 @@ function collectInputSourceReferences(inputMessages) {
       references.push({
         kind: "input_file",
         displayPath,
-        ...typeof segment.resolvedPath === "string" ? { resolvedPath: path45.resolve(segment.resolvedPath) } : {}
+        ...typeof segment.resolvedPath === "string" ? { resolvedPath: path46.resolve(segment.resolvedPath) } : {}
       });
     }
   }
@@ -38134,7 +39054,7 @@ function collectSingleGraderSourceReferences(evaluator) {
       references.push({
         kind: "code_grader_command",
         displayPath: evaluator.aggregator.path,
-        resolvedPath: path45.resolve(evaluator.aggregator.cwd ?? "", evaluator.aggregator.path),
+        resolvedPath: path46.resolve(evaluator.aggregator.cwd ?? "", evaluator.aggregator.path),
         graderName: evaluator.name
       });
     } else if (evaluator.aggregator.type === "llm-grader" && evaluator.aggregator.promptPath) {
@@ -38167,9 +39087,9 @@ function dedupeSourceReferences(references) {
   return deduped;
 }
 function toPortableRelativePath(root, candidate) {
-  const relative = path45.relative(root, candidate);
-  if (relative && !relative.startsWith("..") && !path45.isAbsolute(relative)) {
-    return relative.split(path45.sep).join("/");
+  const relative = path46.relative(root, candidate);
+  if (relative && !relative.startsWith("..") && !path46.isAbsolute(relative)) {
+    return relative.split(path46.sep).join("/");
   }
   return void 0;
 }
@@ -38223,8 +39143,8 @@ function parseWorkspaceScriptConfig(raw, evalFileDir) {
   if (!command) return void 0;
   const timeoutMs = typeof obj.timeout_ms === "number" ? obj.timeout_ms : void 0;
   let cwd = typeof obj.cwd === "string" ? obj.cwd : void 0;
-  if (cwd && !path45.isAbsolute(cwd)) {
-    cwd = path45.resolve(evalFileDir, cwd);
+  if (cwd && !path46.isAbsolute(cwd)) {
+    cwd = path46.resolve(evalFileDir, cwd);
   }
   const config2 = { command };
   if (timeoutMs !== void 0) {
@@ -38262,7 +39182,7 @@ function parseWorkspaceHooksConfig(raw, evalFileDir) {
 }
 async function resolveWorkspaceConfig(raw, evalFileDir) {
   if (typeof raw === "string") {
-    const workspaceFilePath = path45.resolve(evalFileDir, raw);
+    const workspaceFilePath = path46.resolve(evalFileDir, raw);
     let content;
     try {
       content = await readFile18(workspaceFilePath, "utf8");
@@ -38275,7 +39195,7 @@ async function resolveWorkspaceConfig(raw, evalFileDir) {
         `Invalid workspace file format: ${workspaceFilePath} (expected a YAML object)`
       );
     }
-    const workspaceFileDir = path45.dirname(workspaceFilePath);
+    const workspaceFileDir = path46.dirname(workspaceFilePath);
     const resolvedWorkspace = parseWorkspaceConfig(parsed, workspaceFileDir);
     if (resolvedWorkspace) {
       return { ...resolvedWorkspace, workspaceFileDir };
@@ -38309,8 +39229,8 @@ function parseWorkspaceConfig(raw, evalFileDir) {
     throw new Error("workspace.static has been removed. Use workspace.mode='static'.");
   }
   let template = typeof obj.template === "string" ? obj.template : void 0;
-  if (template && !path45.isAbsolute(template)) {
-    template = path45.resolve(evalFileDir, template);
+  if (template && !path46.isAbsolute(template)) {
+    template = path46.resolve(evalFileDir, template);
   }
   const isolation = obj.isolation === "shared" || obj.isolation === "per_test" ? obj.isolation : void 0;
   const repos = Array.isArray(obj.repos) ? obj.repos.map(parseRepoConfig).filter(Boolean) : void 0;
@@ -38447,8 +39367,6 @@ ${detailBlock}${ANSI_RESET6}`);
     console.error(`${ANSI_RED3}Error: ${message}${ANSI_RESET6}`);
   }
 }
-var execFileAsync2 = promisify6(execFile2);
-var WORKSPACE_GIT_TIMEOUT_MS = 3e5;
 function pathFromRoot(root) {
   return root instanceof URL ? fileURLToPath5(root) : String(root);
 }
@@ -38470,53 +39388,84 @@ function buildSkippedEvaluatorError(scores) {
 function usesFileReferencePrompt(provider) {
   return isAgentProvider(provider) || provider.kind === "cli";
 }
-function toScriptConfig(hook, hookName, context) {
-  const command = hook.command ?? hook.script;
-  if (!command || command.length === 0) {
-    throw new Error(`${hookName} hook in ${context} requires command or script`);
+function createEvaluationRuntime(options) {
+  const {
+    target,
+    targets,
+    env,
+    providerFactory,
+    evalFilePath,
+    graderTarget: cliGraderTarget,
+    model: cliModel
+  } = options;
+  const resolvedTargetsByName = /* @__PURE__ */ new Map();
+  resolvedTargetsByName.set(target.name, target);
+  const targetDefinitions = /* @__PURE__ */ new Map();
+  for (const definition of targets ?? []) {
+    targetDefinitions.set(definition.name, definition);
   }
-  return {
-    command,
-    ...hook.timeout_ms !== void 0 && { timeout_ms: hook.timeout_ms },
-    ...hook.timeoutMs !== void 0 && { timeoutMs: hook.timeoutMs },
-    ...hook.cwd !== void 0 && { cwd: hook.cwd },
-    ...hook.script !== void 0 && { script: hook.script }
+  const envLookup = env ?? process.env;
+  const providerCache = /* @__PURE__ */ new Map();
+  const getOrCreateProvider = (resolved) => {
+    const existing = providerCache.get(resolved.name);
+    if (existing) {
+      return existing;
+    }
+    const factory = providerFactory ?? createProvider;
+    const instance = factory(resolved);
+    providerCache.set(resolved.name, instance);
+    return instance;
   };
-}
-function hasHookCommand(hook) {
-  return !!(hook?.command && hook.command.length > 0 || hook?.script && hook.script.length > 0);
-}
-function hooksEnabled(workspace) {
-  return workspace?.hooks?.enabled !== false;
-}
-function workspaceGitEnv() {
-  const env = { ...process.env };
-  for (const key of Object.keys(env)) {
-    if (key.startsWith("GIT_") && key !== "GIT_SSH_COMMAND") {
-      delete env[key];
+  const resolveTargetByName = (name) => {
+    if (resolvedTargetsByName.has(name)) {
+      return resolvedTargetsByName.get(name);
     }
-  }
-  return {
-    ...env,
-    GIT_TERMINAL_PROMPT: "0",
-    GIT_ASKPASS: "",
-    GIT_SSH_COMMAND: "ssh -o BatchMode=yes"
+    const definition = resolveDelegatedTargetDefinition(name, targetDefinitions, envLookup);
+    if (!definition) {
+      return void 0;
+    }
+    const resolved = resolveTargetDefinition(definition, envLookup, evalFilePath ?? "");
+    resolvedTargetsByName.set(name, resolved);
+    return resolved;
   };
-}
-async function resetWorkspaceRoot(workspacePath, resetMode, baselineRef) {
-  if (!existsSync6(path46.join(workspacePath, ".git"))) {
-    return false;
-  }
-  const cleanFlag = resetMode === "strict" ? "-fdx" : "-fd";
-  const opts = {
-    cwd: workspacePath,
-    timeout: WORKSPACE_GIT_TIMEOUT_MS,
-    env: workspaceGitEnv(),
-    maxBuffer: 50 * 1024 * 1024
+  const resolveGraderProvider = async (targetContext) => {
+    if (cliGraderTarget) {
+      if (cliGraderTarget === "agentv") {
+        if (!cliModel) {
+          throw new Error('--grader-target "agentv" requires --model (e.g., "openai:gpt-5-mini")');
+        }
+        const { AgentvProvider: AgentvProvider2 } = await import("./agentv-provider-AYXH7WLW-NJRC6UQX.js");
+        return new AgentvProvider2("agentv", { model: cliModel, temperature: 0 });
+      }
+      const overrideTarget = resolveTargetByName(cliGraderTarget);
+      if (!overrideTarget) {
+        throw new Error(`--grader-target "${cliGraderTarget}" not found in targets`);
+      }
+      return getOrCreateProvider(overrideTarget);
+    }
+    const graderName = targetContext.graderTarget ?? targetContext.name;
+    const resolvedGrader = resolveTargetByName(graderName);
+    if (!resolvedGrader) {
+      if (!LLM_GRADER_CAPABLE_KINDS.includes(targetContext.kind)) {
+        return void 0;
+      }
+      return getOrCreateProvider(targetContext);
+    }
+    return getOrCreateProvider(resolvedGrader);
+  };
+  const targetResolver = (name) => {
+    const resolved = resolveTargetByName(name);
+    if (!resolved) {
+      return void 0;
+    }
+    return getOrCreateProvider(resolved);
+  };
+  return {
+    getOrCreateProvider,
+    resolveGraderProvider,
+    targetResolver,
+    availableTargets: [target.name, ...Array.from(targetDefinitions.keys())]
   };
-  await execFileAsync2("git", ["reset", "--hard", baselineRef ?? "HEAD"], opts);
-  await execFileAsync2("git", ["clean", cleanFlag], opts);
-  return true;
 }
 function validateDependencyGraph(tests) {
   const ids = /* @__PURE__ */ new Set();
@@ -38547,18 +39496,18 @@ function validateDependencyGraph(tests) {
   }
   const visited = /* @__PURE__ */ new Set();
   const visiting = /* @__PURE__ */ new Set();
-  function visit(id, path50) {
+  function visit(id, path51) {
     if (visiting.has(id)) {
-      const cycle = [...path50.slice(path50.indexOf(id)), id];
+      const cycle = [...path51.slice(path51.indexOf(id)), id];
       throw new Error(`Circular dependency detected: ${cycle.join(" \u2192 ")}`);
     }
     if (visited.has(id)) return;
     visiting.add(id);
-    path50.push(id);
+    path51.push(id);
     for (const dep of depMap.get(id) ?? []) {
-      visit(dep, path50);
+      visit(dep, path51);
     }
-    path50.pop();
+    path51.pop();
     visiting.delete(id);
     visited.add(id);
   }
@@ -38612,6 +39561,170 @@ function computeWaves(tests) {
   }
   return waves;
 }
+function createPreparedProvider(target) {
+  return {
+    id: `prepared:${target.name}`,
+    kind: target.kind,
+    targetName: target.name,
+    async invoke() {
+      throw new Error("Prepared grading does not invoke the target provider");
+    }
+  };
+}
+function withPreparedMetadata(evalCase, preparedAttempt) {
+  return {
+    ...evalCase.metadata,
+    preparedAttempt
+  };
+}
+async function gradePreparedEvalCase(options) {
+  const {
+    evalCase,
+    target,
+    targets,
+    env,
+    evaluators,
+    providerFactory,
+    agentTimeoutMs,
+    graderTarget,
+    model,
+    evalFilePath,
+    workspacePath,
+    baselineCommit,
+    response,
+    verbose,
+    threshold: caseThreshold,
+    preparedAttempt
+  } = options;
+  const nowFn = options.now ?? (() => /* @__PURE__ */ new Date());
+  const caseStartMs = Date.now();
+  const provider = createPreparedProvider(target);
+  const formattingMode = usesFileReferencePrompt(provider) ? "agent" : "lm";
+  const promptInputs = await buildPromptInputs(evalCase, formattingMode);
+  const typeRegistry = createBuiltinRegistry();
+  const runtime = createEvaluationRuntime({
+    target,
+    targets,
+    env,
+    providerFactory,
+    evalFilePath,
+    graderTarget,
+    model
+  });
+  const evaluatorRegistry = buildEvaluatorRegistry(evaluators, runtime.resolveGraderProvider);
+  const discoveryBaseDir = evalFilePath ? path47.dirname(path47.resolve(evalFilePath)) : process.cwd();
+  await discoverAssertions(typeRegistry, discoveryBaseDir);
+  await discoverGraders(typeRegistry, discoveryBaseDir);
+  let fileChanges;
+  if (baselineCommit) {
+    try {
+      const diff = await captureFileChanges(workspacePath, baselineCommit);
+      if (diff.length > 0) {
+        fileChanges = diff;
+      }
+    } catch (error40) {
+      if (verbose) {
+        const message = error40 instanceof Error ? error40.message : String(error40);
+        console.warn(`Warning: failed to capture prepared workspace diff: ${message}`);
+      }
+    }
+  }
+  const candidate = response ?? "";
+  const input = buildResultInput(promptInputs);
+  const outputMessages = candidate.length > 0 ? [{ role: "assistant", content: candidate }] : [];
+  const resultTrace = buildTraceFromMessages({
+    input,
+    output: outputMessages,
+    finalOutput: candidate,
+    provider: provider.kind,
+    target: target.name,
+    testId: evalCase.id,
+    conversationId: evalCase.conversation_id
+  });
+  try {
+    const gradeStartedAt = nowFn();
+    const { score, scores } = await runEvaluatorsForCase({
+      evalCase,
+      candidate,
+      target,
+      provider,
+      evaluators: evaluatorRegistry,
+      typeRegistry,
+      attempt: 0,
+      promptInputs,
+      now: gradeStartedAt,
+      agentTimeoutMs,
+      targetResolver: runtime.targetResolver,
+      availableTargets: runtime.availableTargets,
+      fileChanges,
+      workspacePath,
+      dockerConfig: evalCase.workspace?.docker,
+      threshold: evalCase.threshold ?? caseThreshold
+    });
+    const timestamp = nowFn();
+    const effectiveThreshold = evalCase.threshold ?? caseThreshold;
+    const graderTokens = aggregateEvaluatorTokenUsage(scores);
+    const evalRun = {
+      durationMs: Date.now() - caseStartMs,
+      ...graderTokens ? { tokenUsage: graderTokens } : {}
+    };
+    const skippedEvaluatorError = buildSkippedEvaluatorError(scores);
+    const executionStatus = skippedEvaluatorError ? "execution_error" : classifyQualityStatus(score.score, effectiveThreshold);
+    const baseResult = {
+      timestamp: timestamp.toISOString(),
+      testId: evalCase.id,
+      suite: evalCase.suite,
+      category: evalCase.category,
+      conversationId: evalCase.conversation_id,
+      score: skippedEvaluatorError ? 0 : score.score,
+      assertions: score.assertions,
+      target: target.name,
+      input,
+      output: candidate,
+      scores,
+      trace: resultTrace,
+      fileChanges,
+      workspacePath,
+      evalRun,
+      metadata: withPreparedMetadata(evalCase, preparedAttempt),
+      executionStatus
+    };
+    if (!skippedEvaluatorError) {
+      return baseResult;
+    }
+    return {
+      ...baseResult,
+      trace: appendErrorEventToTrace(baseResult.trace, skippedEvaluatorError, {
+        failure_stage: "evaluator",
+        failure_reason_code: "evaluator_error"
+      }),
+      error: skippedEvaluatorError,
+      failureStage: "evaluator",
+      failureReasonCode: "evaluator_error",
+      executionError: { message: skippedEvaluatorError, stage: "evaluator" }
+    };
+  } catch (error40) {
+    const evalRun = { durationMs: Date.now() - caseStartMs };
+    const errorResult = buildErrorResult(
+      evalCase,
+      target.name,
+      nowFn(),
+      error40,
+      promptInputs,
+      provider,
+      "evaluator",
+      "evaluator_error",
+      verbose
+    );
+    return {
+      ...errorResult,
+      evalRun,
+      fileChanges,
+      workspacePath,
+      metadata: withPreparedMetadata(evalCase, preparedAttempt)
+    };
+  }
+}
 async function runEvaluation(options) {
   const {
     testFilePath: evalFilePath,
@@ -38667,80 +39780,24 @@ async function runEvaluation(options) {
     }
     return [];
   }
-  const resolvedTargetsByName = /* @__PURE__ */ new Map();
-  resolvedTargetsByName.set(target.name, target);
-  const targetDefinitions = /* @__PURE__ */ new Map();
-  for (const definition of targets ?? []) {
-    targetDefinitions.set(definition.name, definition);
-  }
-  const envLookup = env ?? process.env;
-  const providerCache = /* @__PURE__ */ new Map();
-  const getOrCreateProvider = (resolved) => {
-    const existing = providerCache.get(resolved.name);
-    if (existing) {
-      return existing;
-    }
-    const factory = providerFactory ?? createProvider;
-    const instance = factory(resolved);
-    providerCache.set(resolved.name, instance);
-    return instance;
-  };
-  const resolveTargetByName = (name) => {
-    if (resolvedTargetsByName.has(name)) {
-      return resolvedTargetsByName.get(name);
-    }
-    const definition = resolveDelegatedTargetDefinition(name, targetDefinitions, envLookup);
-    if (!definition) {
-      return void 0;
-    }
-    const resolved = resolveTargetDefinition(definition, envLookup, evalFilePath);
-    resolvedTargetsByName.set(name, resolved);
-    return resolved;
-  };
-  const resolveGraderProvider = async (targetContext) => {
-    if (cliGraderTarget) {
-      if (cliGraderTarget === "agentv") {
-        if (!cliModel) {
-          throw new Error('--grader-target "agentv" requires --model (e.g., "openai:gpt-5-mini")');
-        }
-        const { AgentvProvider: AgentvProvider2 } = await import("./agentv-provider-AYXH7WLW-NJRC6UQX.js");
-        return new AgentvProvider2("agentv", { model: cliModel, temperature: 0 });
-      }
-      const overrideTarget = resolveTargetByName(cliGraderTarget);
-      if (!overrideTarget) {
-        throw new Error(`--grader-target "${cliGraderTarget}" not found in targets`);
-      }
-      return getOrCreateProvider(overrideTarget);
-    }
-    const graderName = targetContext.graderTarget ?? targetContext.name;
-    const resolvedGrader = resolveTargetByName(graderName);
-    if (!resolvedGrader) {
-      if (!LLM_GRADER_CAPABLE_KINDS.includes(targetContext.kind)) {
-        return void 0;
-      }
-      return getOrCreateProvider(targetContext);
-    }
-    return getOrCreateProvider(resolvedGrader);
-  };
+  const runtime = createEvaluationRuntime({
+    target,
+    targets,
+    env,
+    providerFactory,
+    evalFilePath,
+    graderTarget: cliGraderTarget,
+    model: cliModel
+  });
+  const { getOrCreateProvider, resolveGraderProvider, targetResolver, availableTargets } = runtime;
   if (isAgentProvider(getOrCreateProvider(target)) && !target.graderTarget && !cliGraderTarget) {
     throw new Error(
       `Target "${target.name}" is an agent provider ("${target.kind}") with no grader_target \u2014 agent providers cannot return structured JSON for grading. Set grader_target to an LLM provider (e.g., azure-llm).`
     );
   }
-  const targetResolver = (name) => {
-    const resolved = resolveTargetByName(name);
-    if (!resolved) {
-      return void 0;
-    }
-    return getOrCreateProvider(resolved);
-  };
-  const availableTargets = [
-    target.name,
-    ...Array.from(targetDefinitions.keys())
-  ];
   const evaluatorRegistry = buildEvaluatorRegistry(evaluators, resolveGraderProvider);
   const typeRegistry = createBuiltinRegistry();
-  const discoveryBaseDir = evalFilePath ? path46.dirname(path46.resolve(evalFilePath)) : process.cwd();
+  const discoveryBaseDir = evalFilePath ? path47.dirname(path47.resolve(evalFilePath)) : process.cwd();
   const evalDir = discoveryBaseDir;
   await discoverAssertions(typeRegistry, discoveryBaseDir);
   await discoverGraders(typeRegistry, discoveryBaseDir);
@@ -38796,132 +39853,38 @@ async function runEvaluation(options) {
       }
     }
   }
-  const suiteWorkspace = filteredEvalCases[0]?.workspace;
-  const rawTemplate = suiteWorkspace?.template;
-  const resolvedTemplate = await resolveWorkspaceTemplate(rawTemplate);
-  const workspaceTemplate = resolvedTemplate?.dir;
-  let suiteWorkspaceFile = resolvedTemplate?.workspaceFile;
-  const setupLog = (message) => {
-    if (verbose) {
-      console.log(`[setup] ${message}`);
-    }
-  };
-  const isPerTestIsolation = suiteWorkspace?.isolation === "per_test";
-  const cliWorkspacePath = workspacePath ?? legacyWorkspacePath;
-  const yamlWorkspacePath = suiteWorkspace?.path;
-  if (cliWorkspacePath && workspaceMode && workspaceMode !== "static") {
-    throw new Error("--workspace-path requires --workspace-mode static when both are provided");
-  }
-  let configuredMode = cliWorkspacePath ? "static" : workspaceMode ?? suiteWorkspace?.mode ?? (yamlWorkspacePath ? "static" : "pooled");
-  const configuredStaticPath = cliWorkspacePath ?? yamlWorkspacePath;
-  if (configuredMode === "static" && !configuredStaticPath) {
-    if (!suiteWorkspace?.repos?.length) {
-      setupLog("workspace.mode=static with no path and no repos \u2014 falling back to temp mode");
-      configuredMode = "temp";
-    } else {
-      throw new Error("workspace.mode=static requires workspace.path or --workspace-path");
-    }
-  }
-  const useStaticWorkspace = configuredMode === "static";
-  if (useStaticWorkspace && isPerTestIsolation) {
-    throw new Error(
-      "static workspace mode is incompatible with isolation: per_test. Use isolation: shared (default)."
-    );
-  }
-  if (configuredMode !== "static" && configuredStaticPath) {
-    throw new Error("workspace.path requires workspace.mode=static");
-  }
-  const hasSharedWorkspace = !!(useStaticWorkspace || !isPerTestIsolation && (workspaceTemplate || suiteWorkspace?.hooks || suiteWorkspace?.repos?.length));
-  const poolEnabled = configuredMode === "pooled";
-  const usePool = poolEnabled !== false && !!suiteWorkspace?.repos?.length && !isPerTestIsolation && !useStaticWorkspace;
   const resolvedRetainOnSuccess = retainOnSuccess ?? (keepWorkspaces ? "keep" : "cleanup");
   const resolvedRetainOnFailure = retainOnFailure ?? (cleanupWorkspaces ? "cleanup" : "keep");
   const workers = options.maxConcurrency ?? target.workers ?? 1;
-  setupLog(
-    `sharedWorkspace=${hasSharedWorkspace} perTestIsolation=${isPerTestIsolation} usePool=${usePool} workers=${workers}`
-  );
-  if (hasSharedWorkspace && !usePool && workers > 1 && filteredEvalCases.length > 1) {
-    console.warn(
-      [
-        `Warning: This eval uses a shared workspace with ${workers} workers.`,
-        "If the agent under test makes file edits, concurrent runs may corrupt each other.",
-        "To limit concurrency, add this to your eval YAML:",
-        "",
-        "  execution:",
-        "    workers: 1",
-        "",
-        "Or pass --workers 1 on the command line."
-      ].join("\n")
-    );
-  }
   const limit = pLimit(workers);
-  let sharedWorkspacePath;
-  let sharedBaselineCommit;
-  let beforeAllOutput;
-  let poolManager;
-  let poolSlot;
-  const poolSlots = [];
-  const availablePoolSlots = [];
-  const poolSlotBaselines = /* @__PURE__ */ new Map();
-  const poolMaxSlots = Math.min(configPoolMaxSlots ?? 10, 50);
-  let staticMaterialised = false;
-  const isYamlConfiguredPath = !cliWorkspacePath && !!yamlWorkspacePath;
-  if (useStaticWorkspace && configuredStaticPath) {
-    const dirExists = await stat9(configuredStaticPath).then(
-      (s) => s.isDirectory(),
-      () => false
-    );
-    const isEmpty = dirExists ? (await readdir8(configuredStaticPath)).length === 0 : false;
-    if (isYamlConfiguredPath && (!dirExists || isEmpty)) {
-      if (!dirExists) {
-        await mkdir17(configuredStaticPath, { recursive: true });
-      }
-      if (workspaceTemplate) {
-        await copyDirectoryRecursive(workspaceTemplate, configuredStaticPath);
-        setupLog(`copied template into static workspace: ${configuredStaticPath}`);
-      }
-      staticMaterialised = true;
-      setupLog(`materialised static workspace at: ${configuredStaticPath}`);
-    } else {
-      setupLog(`reusing existing static workspace: ${configuredStaticPath}`);
-    }
-    sharedWorkspacePath = configuredStaticPath;
-  } else if (!isPerTestIsolation && usePool && suiteWorkspace?.repos) {
-    const slotsNeeded = workers;
-    setupLog(`acquiring ${slotsNeeded} workspace pool slot(s) (pool capacity: ${poolMaxSlots})`);
-    poolManager = new WorkspacePoolManager(getWorkspacePoolRoot());
-    const poolRepoManager = new RepoManager(verbose);
-    for (let i = 0; i < slotsNeeded; i++) {
-      const slot = await poolManager.acquireWorkspace({
-        templatePath: workspaceTemplate,
-        repos: suiteWorkspace.repos,
-        maxSlots: poolMaxSlots,
-        repoManager: poolRepoManager,
-        poolReset: (workspaceClean === "full" ? "strict" : workspaceClean === "standard" ? "fast" : null) ?? "fast"
-      });
-      poolSlots.push(slot);
-      setupLog(`pool slot ${i} acquired at: ${slot.path} (existing=${slot.isExisting})`);
-    }
-    if (slotsNeeded === 1) {
-      poolSlot = poolSlots[0];
-      sharedWorkspacePath = poolSlot.path;
-    } else {
-      availablePoolSlots.push(...poolSlots);
-    }
-  } else if (!isPerTestIsolation && workspaceTemplate) {
-    setupLog(`creating shared workspace from template: ${workspaceTemplate}`);
-    try {
-      sharedWorkspacePath = await createTempWorkspace(workspaceTemplate, evalRunId, "shared");
-      setupLog(`shared workspace created at: ${sharedWorkspacePath}`);
-    } catch (error40) {
-      const message = error40 instanceof Error ? error40.message : String(error40);
-      throw new Error(`Failed to create shared workspace: ${message}`);
-    }
-  } else if (!isPerTestIsolation && (suiteWorkspace?.hooks || suiteWorkspace?.repos?.length)) {
-    sharedWorkspacePath = getWorkspacePath(evalRunId, "shared");
-    await mkdir17(sharedWorkspacePath, { recursive: true });
-    setupLog(`created empty shared workspace at: ${sharedWorkspacePath}`);
-  }
+  const sharedSetup = await prepareSharedWorkspaceSetup({
+    evalRunId,
+    evalCases: filteredEvalCases,
+    targetHooks: options.targetHooks,
+    evalDir,
+    verbose,
+    workers,
+    poolMaxSlots: configPoolMaxSlots,
+    workspacePath,
+    legacyWorkspacePath,
+    workspaceMode,
+    workspaceClean
+  });
+  const {
+    suiteWorkspace,
+    sharedWorkspacePath,
+    sharedBaselineCommit,
+    suiteWorkspaceFile,
+    beforeAllOutput,
+    repoManager,
+    poolSlot,
+    poolSlots,
+    availablePoolSlots,
+    poolSlotBaselines,
+    useStaticWorkspace
+  } = sharedSetup;
+  const targetHooks = options.targetHooks;
+  const suiteHooksEnabled = hooksEnabled(suiteWorkspace);
   try {
     let toDependencyResult2 = function(r) {
       return {
@@ -38959,198 +39922,6 @@ async function runEvaluation(options) {
       return result.costUsd;
     };
     var toDependencyResult = toDependencyResult2, checkDependencies = checkDependencies2, extractEvaluationCostUsd = extractEvaluationCostUsd2;
-    if (suiteWorkspaceFile && sharedWorkspacePath) {
-      const copiedWorkspaceFile = path46.join(sharedWorkspacePath, path46.basename(suiteWorkspaceFile));
-      try {
-        await stat9(copiedWorkspaceFile);
-        suiteWorkspaceFile = copiedWorkspaceFile;
-      } catch {
-      }
-    }
-    const hasReposToMaterialize = !!suiteWorkspace?.repos?.length && !usePool && !isPerTestIsolation;
-    const needsRepoMaterialisation = hasReposToMaterialize && (!useStaticWorkspace || staticMaterialised);
-    const needsPerRepoCheck = hasReposToMaterialize && useStaticWorkspace && !staticMaterialised && isYamlConfiguredPath;
-    const repoManager = needsRepoMaterialisation || needsPerRepoCheck ? new RepoManager(verbose) : void 0;
-    if (repoManager && sharedWorkspacePath && suiteWorkspace?.repos) {
-      try {
-        if (needsPerRepoCheck) {
-          for (const repo of suiteWorkspace.repos) {
-            if (!repo.path || !repo.repo) continue;
-            const targetDir = path46.join(sharedWorkspacePath, repo.path);
-            if (existsSync6(targetDir)) {
-              setupLog(`reusing existing repo at: ${targetDir}`);
-              continue;
-            }
-            setupLog(`materializing missing repo: ${repo.path}`);
-            await repoManager.materialize(repo, sharedWorkspacePath);
-          }
-        } else {
-          setupLog(
-            `materializing ${suiteWorkspace.repos.length} shared repo(s) into ${sharedWorkspacePath}`
-          );
-          await repoManager.materializeAll(suiteWorkspace.repos, sharedWorkspacePath);
-        }
-        setupLog("shared repo materialization complete");
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        if (sharedWorkspacePath && !useStaticWorkspace) {
-          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-          });
-        }
-        throw new Error(`Failed to materialize repos: ${message}`);
-      }
-    }
-    const suiteDockerConfig = suiteWorkspace?.docker;
-    if (suiteDockerConfig) {
-      setupLog(`pulling Docker image: ${suiteDockerConfig.image}`);
-      const { DockerWorkspaceProvider } = await import("./docker-workspace-RPPXBT27-B4AQHVWA.js");
-      const dockerSetup = new DockerWorkspaceProvider(suiteDockerConfig);
-      if (!await dockerSetup.isDockerAvailable()) {
-        throw new Error(
-          "Docker workspace configured but Docker CLI is not available. Install Docker and ensure it is running."
-        );
-      }
-      await dockerSetup.pullImage();
-      setupLog("Docker image pull complete");
-    }
-    if (suiteWorkspace?.env) {
-      try {
-        await runPreflightChecks(suiteWorkspace.env, sharedWorkspacePath ?? void 0, setupLog);
-        setupLog("preflight checks passed");
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        if (sharedWorkspacePath && !useStaticWorkspace) {
-          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-          });
-        }
-        throw new Error(message);
-      }
-    }
-    const suiteHooksEnabled = hooksEnabled(suiteWorkspace);
-    const suiteBeforeAllHook = suiteWorkspace?.hooks?.before_all;
-    if (sharedWorkspacePath && suiteHooksEnabled && hasHookCommand(suiteBeforeAllHook)) {
-      const beforeAllHook = suiteBeforeAllHook;
-      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
-      setupLog(
-        `running shared before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
-      );
-      const scriptContext = {
-        workspacePath: sharedWorkspacePath,
-        testId: "__before_all__",
-        evalRunId,
-        evalDir,
-        workspaceFileDir: suiteWorkspace?.workspaceFileDir
-      };
-      try {
-        beforeAllOutput = await executeWorkspaceScript(
-          toScriptConfig(beforeAllHook, "before_all", "suite workspace"),
-          scriptContext
-        );
-        setupLog("shared before_all completed");
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        if (sharedWorkspacePath && !useStaticWorkspace) {
-          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-          });
-        }
-        throw new Error(`before_all script failed: ${message}`);
-      }
-    }
-    if (availablePoolSlots.length > 0 && suiteHooksEnabled && hasHookCommand(suiteBeforeAllHook)) {
-      const beforeAllHook = suiteBeforeAllHook;
-      for (const slot of availablePoolSlots) {
-        setupLog(`running before_all on pool slot ${slot.index}`);
-        const scriptContext = {
-          workspacePath: slot.path,
-          testId: "__before_all__",
-          evalRunId,
-          evalDir,
-          workspaceFileDir: suiteWorkspace?.workspaceFileDir
-        };
-        try {
-          const output = await executeWorkspaceScript(
-            toScriptConfig(beforeAllHook, "before_all", "suite workspace"),
-            scriptContext
-          );
-          if (!beforeAllOutput) beforeAllOutput = output;
-          setupLog(`before_all completed on pool slot ${slot.index}`);
-        } catch (error40) {
-          const message = error40 instanceof Error ? error40.message : String(error40);
-          throw new Error(`before_all script failed on pool slot ${slot.index}: ${message}`);
-        }
-      }
-    }
-    const targetHooks = options.targetHooks;
-    const targetBeforeAllHook = targetHooks?.before_all;
-    if (sharedWorkspacePath && hasHookCommand(targetBeforeAllHook)) {
-      const beforeAllCommand = (targetBeforeAllHook.command ?? []).join(" ");
-      setupLog(`running target before_all command=${beforeAllCommand}`);
-      const scriptContext = {
-        workspacePath: sharedWorkspacePath,
-        testId: "__target_before_all__",
-        evalRunId,
-        evalDir,
-        workspaceFileDir: suiteWorkspace?.workspaceFileDir
-      };
-      try {
-        await executeWorkspaceScript(
-          toScriptConfig(targetBeforeAllHook, "before_all", "target hooks"),
-          scriptContext
-        );
-        setupLog("target before_all completed");
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        if (sharedWorkspacePath && !useStaticWorkspace) {
-          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-          });
-        }
-        throw new Error(`target before_all hook failed: ${message}`);
-      }
-    }
-    if (availablePoolSlots.length > 0 && hasHookCommand(targetBeforeAllHook)) {
-      for (const slot of availablePoolSlots) {
-        setupLog(`running target before_all on pool slot ${slot.index}`);
-        const scriptContext = {
-          workspacePath: slot.path,
-          testId: "__target_before_all__",
-          evalRunId,
-          evalDir,
-          workspaceFileDir: suiteWorkspace?.workspaceFileDir
-        };
-        try {
-          await executeWorkspaceScript(
-            toScriptConfig(targetBeforeAllHook, "before_all", "target hooks"),
-            scriptContext
-          );
-        } catch (error40) {
-          const message = error40 instanceof Error ? error40.message : String(error40);
-          throw new Error(`target before_all hook failed on pool slot ${slot.index}: ${message}`);
-        }
-      }
-    }
-    if (sharedWorkspacePath) {
-      try {
-        sharedBaselineCommit = await initializeBaseline(sharedWorkspacePath);
-        setupLog(`shared baseline initialized: ${sharedBaselineCommit}`);
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        setupLog(`shared baseline initialization failed (file_changes unavailable): ${message}`);
-      }
-    }
-    if (availablePoolSlots.length > 0) {
-      for (const slot of availablePoolSlots) {
-        try {
-          const baseline = await initializeBaseline(slot.path);
-          poolSlotBaselines.set(slot.path, baseline);
-          setupLog(`pool slot ${slot.index} baseline initialized: ${baseline}`);
-        } catch (error40) {
-          const message = error40 instanceof Error ? error40.message : String(error40);
-          setupLog(
-            `pool slot ${slot.index} baseline initialization failed (file_changes unavailable): ${message}`
-          );
-        }
-      }
-    }
     let nextWorkerId = 1;
     const workerIdByEvalId = /* @__PURE__ */ new Map();
     let beforeAllOutputAttached = false;
@@ -39554,17 +40325,7 @@ async function runEvaluation(options) {
     }
     return results;
   } finally {
-    if (poolManager) {
-      if (poolSlot) {
-        await poolManager.releaseSlot(poolSlot);
-      }
-      for (const slot of poolSlots) {
-        if (slot !== poolSlot) {
-          await poolManager.releaseSlot(slot).catch(() => {
-          });
-        }
-      }
-    }
+    await releaseSharedWorkspaceSetup(sharedSetup);
   }
 }
 async function runBatchEvaluation(options) {
@@ -39812,257 +40573,45 @@ async function runEvalCase(options) {
     cachedResponse = await cache.get(cacheKey);
   }
   const nowFn = now ?? (() => /* @__PURE__ */ new Date());
-  let workspacePath = sharedWorkspacePath;
-  let beforeAllOutput;
-  let beforeEachOutput;
   let afterEachOutput;
-  const isSharedWorkspace = !!sharedWorkspacePath;
-  let caseWorkspaceFile;
   const caseHooksEnabled = hooksEnabled(evalCase.workspace);
-  if (!workspacePath) {
-    const rawCaseTemplate = evalCase.workspace?.template;
-    const resolvedCaseTemplate = await resolveWorkspaceTemplate(rawCaseTemplate);
-    const caseWorkspaceTemplate = resolvedCaseTemplate?.dir;
-    caseWorkspaceFile = resolvedCaseTemplate?.workspaceFile;
-    if (caseWorkspaceTemplate && evalRunId) {
-      try {
-        workspacePath = await createTempWorkspace(caseWorkspaceTemplate, evalRunId, evalCase.id);
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        return buildErrorResult(
-          evalCase,
-          target.name,
-          nowFn(),
-          new Error(`Failed to create workspace: ${message}`),
-          promptInputs,
-          provider,
-          "setup",
-          "template_error",
-          verbose
-        );
-      }
-      if (caseWorkspaceFile && workspacePath) {
-        const copiedFile = path46.join(workspacePath, path46.basename(caseWorkspaceFile));
-        try {
-          await stat9(copiedFile);
-          caseWorkspaceFile = copiedFile;
-        } catch {
-        }
-      }
-    }
-    if (!workspacePath && (evalCase.workspace?.hooks || evalCase.workspace?.repos?.length) && evalRunId) {
-      workspacePath = getWorkspacePath(evalRunId, evalCase.id);
-      await mkdir17(workspacePath, { recursive: true });
-    }
-    if (evalCase.workspace?.repos?.length && workspacePath) {
-      const perCaseRepoManager = new RepoManager(setupDebug);
-      try {
-        if (setupDebug) {
-          console.log(
-            `[setup] test=${evalCase.id} materializing ${evalCase.workspace.repos.length} per-test repo(s) into ${workspacePath}`
-          );
-        }
-        await perCaseRepoManager.materializeAll(evalCase.workspace.repos, workspacePath);
-        if (setupDebug) {
-          console.log(`[setup] test=${evalCase.id} per-test repo materialization complete`);
-        }
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        return buildErrorResult(
-          evalCase,
-          target.name,
-          nowFn(),
-          new Error(`Failed to materialize repos: ${message}`),
-          promptInputs,
-          provider,
-          "repo_setup",
-          "clone_error",
-          verbose
-        );
-      }
-    }
-    if (workspacePath && evalCase.metadata?.agent_skills_files) {
-      const baseDir = evalCase.metadata.agent_skills_base_dir;
-      const files = evalCase.metadata.agent_skills_files;
-      if (baseDir && files.length > 0) {
-        for (const relPath of files) {
-          const srcPath = path46.resolve(baseDir, relPath);
-          const destPath = path46.resolve(workspacePath, relPath);
-          try {
-            await mkdir17(path46.dirname(destPath), { recursive: true });
-            await copyFile2(srcPath, destPath);
-          } catch (error40) {
-            const message = error40 instanceof Error ? error40.message : String(error40);
-            return buildErrorResult(
-              evalCase,
-              target.name,
-              nowFn(),
-              new Error(
-                `Agent Skills eval file not found: ${relPath} (resolved from ${baseDir}): ${message}`
-              ),
-              promptInputs,
-              provider,
-              "setup",
-              "file_copy_error",
-              verbose
-            );
-          }
-        }
-      }
-    }
-    const caseBeforeAllHook = evalCase.workspace?.hooks?.before_all;
-    if (workspacePath && caseHooksEnabled && hasHookCommand(caseBeforeAllHook)) {
-      const beforeAllHook = caseBeforeAllHook;
-      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
-      if (setupDebug) {
-        console.log(
-          `[setup] test=${evalCase.id} running before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
-        );
-      }
-      const scriptContext = {
-        workspacePath,
-        testId: evalCase.id,
-        evalRunId: evalRunId ?? "",
-        caseInput: evalCase.question,
-        caseMetadata: evalCase.metadata,
-        evalDir,
-        workspaceFileDir: evalCase.workspace?.workspaceFileDir
-      };
-      try {
-        beforeAllOutput = await executeWorkspaceScript(
-          toScriptConfig(beforeAllHook, "before_all", `test '${evalCase.id}'`),
-          scriptContext
-        );
-        if (setupDebug) {
-          console.log(`[setup] test=${evalCase.id} before_all completed`);
-        }
-      } catch (error40) {
-        const message = error40 instanceof Error ? error40.message : String(error40);
-        if (forceCleanup && workspacePath) {
-          await cleanupWorkspace(workspacePath).catch(() => {
-          });
-        }
-        return buildErrorResult(
-          evalCase,
-          target.name,
-          nowFn(),
-          new Error(`before_all script failed: ${message}`),
-          promptInputs,
-          provider,
-          "setup",
-          "script_error",
-          verbose
-        );
-      }
-    }
-  }
-  let beforeEachNeedsFreshBaseline = false;
-  if (caseHooksEnabled && workspacePath && evalCase.workspace?.hooks?.before_each?.reset && evalCase.workspace.hooks.before_each.reset !== "none") {
-    try {
-      if (repoManager && evalCase.workspace.repos?.length) {
-        await repoManager.reset(
-          evalCase.workspace.repos,
-          workspacePath,
-          evalCase.workspace.hooks.before_each.reset
-        );
-      } else {
-        await resetWorkspaceRoot(
-          workspacePath,
-          evalCase.workspace.hooks.before_each.reset,
-          sharedBaselineCommit
-        );
-      }
-    } catch (error40) {
-      const message = error40 instanceof Error ? error40.message : String(error40);
-      return buildErrorResult(
-        evalCase,
-        target.name,
-        nowFn(),
-        new Error(`before_each reset failed: ${message}`),
-        promptInputs,
-        provider,
-        "setup",
-        "script_error",
-        verbose
-      );
-    }
-  }
-  const caseBeforeEachHook = evalCase.workspace?.hooks?.before_each;
-  if (workspacePath && caseHooksEnabled && hasHookCommand(caseBeforeEachHook)) {
-    const beforeEachHook = caseBeforeEachHook;
-    const scriptContext = {
-      workspacePath,
-      testId: evalCase.id,
-      evalRunId: evalRunId ?? "",
-      caseInput: evalCase.question,
-      caseMetadata: evalCase.metadata,
-      evalDir,
-      workspaceFileDir: evalCase.workspace?.workspaceFileDir
-    };
-    try {
-      beforeEachOutput = await executeWorkspaceScript(
-        toScriptConfig(beforeEachHook, "before_each", `test '${evalCase.id}'`),
-        scriptContext
-      );
-      beforeEachNeedsFreshBaseline = true;
-    } catch (error40) {
-      const message = error40 instanceof Error ? error40.message : String(error40);
-      return buildErrorResult(
-        evalCase,
-        target.name,
-        nowFn(),
-        new Error(`before_each script failed: ${message}`),
-        promptInputs,
-        provider,
-        "setup",
-        "script_error",
-        verbose
-      );
-    }
-  }
-  const targetBeforeEachHook = options.targetHooks?.before_each;
-  if (workspacePath && hasHookCommand(targetBeforeEachHook)) {
-    const scriptContext = {
-      workspacePath,
-      testId: evalCase.id,
-      evalRunId: evalRunId ?? "",
-      caseInput: evalCase.question,
-      caseMetadata: evalCase.metadata,
+  let workspaceSetup;
+  try {
+    workspaceSetup = await prepareEvalCaseWorkspace({
+      evalCase,
+      targetName: target.name,
+      evalRunId,
+      sharedWorkspacePath,
+      sharedBaselineCommit,
+      suiteWorkspaceFile,
+      repoManager,
       evalDir,
-      workspaceFileDir: evalCase.workspace?.workspaceFileDir
-    };
-    try {
-      await executeWorkspaceScript(
-        toScriptConfig(targetBeforeEachHook, "before_each", `target hook for '${evalCase.id}'`),
-        scriptContext
-      );
-      beforeEachNeedsFreshBaseline = true;
-    } catch (error40) {
-      const message = error40 instanceof Error ? error40.message : String(error40);
-      return buildErrorResult(
-        evalCase,
-        target.name,
-        nowFn(),
-        new Error(`target before_each hook failed: ${message}`),
-        promptInputs,
-        provider,
-        "setup",
-        "script_error",
-        verbose
-      );
-    }
-  }
-  let baselineCommit = beforeEachNeedsFreshBaseline ? void 0 : sharedBaselineCommit;
-  if (!baselineCommit && workspacePath) {
-    try {
-      baselineCommit = await initializeBaseline(workspacePath);
-    } catch (error40) {
-      const message = error40 instanceof Error ? error40.message : String(error40);
-      if (verbose) {
-        console.warn(`[setup] test=${evalCase.id} baseline initialization failed: ${message}`);
-      }
-    }
+      cleanupWorkspaces: forceCleanup,
+      targetHooks: options.targetHooks,
+      setupDebug
+    });
+  } catch (error40) {
+    const setupError = error40 instanceof WorkspaceSetupError ? error40 : void 0;
+    return buildErrorResult(
+      evalCase,
+      target.name,
+      nowFn(),
+      error40,
+      promptInputs,
+      provider,
+      setupError?.failureStage ?? "setup",
+      setupError?.failureReasonCode ?? "script_error",
+      verbose
+    );
   }
+  const {
+    workspacePath,
+    beforeAllOutput,
+    beforeEachOutput,
+    baselineCommit,
+    isSharedWorkspace,
+    caseWorkspaceFile
+  } = workspaceSetup;
   if (evalCase.mode === "conversation" && evalCase.turns?.length) {
     const conversationResult = await runConversationMode({
       evalCase,
@@ -40769,7 +41318,7 @@ async function runEvaluatorList(options) {
     dockerConfig,
     dependencyResults
   };
-  const evalFileDir = evalCase.file_paths[0] ? path46.dirname(evalCase.file_paths[0]) : process.cwd();
+  const evalFileDir = evalCase.file_paths[0] ? path47.dirname(evalCase.file_paths[0]) : process.cwd();
   const dispatchContext = {
     graderProvider,
     targetResolver,
@@ -41431,38 +41980,6 @@ function computeWeightedMean(entries) {
   }
   return totalWeight > 0 ? weightedSum / totalWeight : 0;
 }
-async function runPreflightChecks(env, cwd, log) {
-  const execFileAsync3 = promisify6(execFile2);
-  const missing = [];
-  for (const cmd of env.required_commands ?? []) {
-    log(`preflight: checking command "${cmd}"`);
-    try {
-      if (process.platform === "win32") {
-        await execFileAsync3("where", [cmd], { cwd });
-      } else {
-        await execFileAsync3("sh", ["-c", `command -v ${cmd}`], { cwd });
-      }
-    } catch {
-      missing.push(`command: ${cmd}`);
-    }
-  }
-  for (const mod of env.required_python_modules ?? []) {
-    log(`preflight: checking Python module "${mod}"`);
-    try {
-      await execFileAsync3("python3", ["-c", `import ${mod}`], { cwd });
-    } catch {
-      missing.push(`python module: ${mod}`);
-    }
-  }
-  if (missing.length > 0) {
-    throw new Error(
-      `Preflight checks failed \u2014 missing dependencies:
-${missing.map((m) => `  \u2022 ${m}`).join("\n")}
-
-Install the missing dependencies before running this eval.`
-    );
-  }
-}
 function createFunctionProvider(taskFn) {
   return {
     id: "function-provider",
@@ -41837,22 +42354,22 @@ function deduplicateByTestIdTarget(results) {
   return deduped;
 }
 async function aggregateRunDir(runDir, options) {
-  const indexPath = path47.join(runDir, RESULT_INDEX_FILENAME);
+  const indexPath = path48.join(runDir, RESULT_INDEX_FILENAME);
   const content = await readFile20(indexPath, "utf8");
   const allResults = parseJsonlResults(content);
   const results = deduplicateByTestIdTarget(allResults);
   const timing = buildTimingArtifact(results);
-  const timingPath = path47.join(runDir, "timing.json");
+  const timingPath = path48.join(runDir, "timing.json");
   await writeFile10(timingPath, `${JSON.stringify(timing, null, 2)}
 `, "utf8");
-  const plannedTestCount = options?.plannedTestCount ?? await readPlannedTestCount(path47.join(runDir, "benchmark.json"));
+  const plannedTestCount = options?.plannedTestCount ?? await readPlannedTestCount(path48.join(runDir, "benchmark.json"));
   const benchmark = buildBenchmarkArtifact(
     results,
     options?.evalFile,
     options?.experiment,
     plannedTestCount
   );
-  const benchmarkPath = path47.join(runDir, "benchmark.json");
+  const benchmarkPath = path48.join(runDir, "benchmark.json");
   await writeFile10(benchmarkPath, `${JSON.stringify(benchmark, null, 2)}
 `, "utf8");
   const targetSet = new Set(results.map((r) => r.target ?? "unknown"));
@@ -41991,17 +42508,37 @@ function toIndexRerunSource(value) {
     source_timestamp: value.sourceTimestamp
   });
 }
+function toIndexPreparedAttempt(value) {
+  if (!isRecord4(value)) {
+    return void 0;
+  }
+  return dropUndefined5({
+    source: value.source,
+    manifest_path: value.manifestPath,
+    prepared_dir: value.preparedDir,
+    workspace_path: value.workspacePath,
+    prompt_path: value.promptPath,
+    target: value.target,
+    prepared_at: value.preparedAt,
+    setup_status: value.setupStatus,
+    baseline_status: value.baselineStatus,
+    baseline_commit: value.baselineCommit
+  });
+}
 function toIndexMetadata(metadata) {
   if (!metadata) {
     return void 0;
   }
   const rerunSource = toIndexRerunSource(metadata.rerunSource);
-  if (!rerunSource) {
+  const preparedAttempt = toIndexPreparedAttempt(metadata.preparedAttempt);
+  if (!rerunSource && !preparedAttempt) {
     return { ...metadata };
   }
+  const reservedKeys = /* @__PURE__ */ new Set(["rerunSource", "preparedAttempt"]);
   return {
-    ...Object.fromEntries(Object.entries(metadata).filter(([key]) => key !== "rerunSource")),
-    rerun_source: rerunSource
+    ...Object.fromEntries(Object.entries(metadata).filter(([key]) => !reservedKeys.has(key))),
+    ...rerunSource ? { rerun_source: rerunSource } : {},
+    ...preparedAttempt ? { prepared_attempt: preparedAttempt } : {}
   };
 }
 function buildGradingArtifact(result) {
@@ -42147,7 +42684,7 @@ async function writeInitialBenchmarkArtifact(runDir, options) {
     options.experiment,
     options.plannedTestCount
   );
-  const benchmarkPath = path47.join(runDir, "benchmark.json");
+  const benchmarkPath = path48.join(runDir, "benchmark.json");
   await writeFile10(benchmarkPath, `${JSON.stringify(stub, null, 2)}
 `, "utf8");
 }
@@ -42197,7 +42734,7 @@ function buildArtifactSubdir(result) {
     segments.push(safeArtifactPathSegment(evalSet, "default"));
   }
   segments.push(safeTestId(result.testId));
-  return path47.posix.join(...segments);
+  return path48.posix.join(...segments);
 }
 function formatOutputMarkdown(output) {
   return output.map((msg) => `@[${msg.role}]:
@@ -42213,7 +42750,7 @@ function extractInput(result) {
   return null;
 }
 function toRelativeArtifactPath(outputDir, filePath) {
-  return path47.relative(outputDir, filePath).split(path47.sep).join("/");
+  return path48.relative(outputDir, filePath).split(path48.sep).join("/");
 }
 function findResultSourceTest(result, testByTestId) {
   return testByTestId.get(result.testId ?? "unknown");
@@ -42229,7 +42766,7 @@ async function writeTraceEnvelopeSidecar(params) {
   const hasTranscript = resultHasExecutionTraceTranscript(params.result);
   const envelope = buildTraceEnvelopeFromEvaluationResult(params.result, {
     evalPath: params.evalPath,
-    runId: path47.basename(params.outputDir),
+    runId: path48.basename(params.outputDir),
     experiment: params.experiment,
     source: { path: RESULT_INDEX_FILENAME },
     capture: { content: "full", redactionLevel: "none", redactedFields: [] },
@@ -42241,7 +42778,7 @@ async function writeTraceEnvelopeSidecar(params) {
     }
   });
   await writeFile10(
-    path47.join(params.outputsDir, "execution-trace.json"),
+    path48.join(params.outputsDir, "execution-trace.json"),
     `${JSON.stringify(toTraceEnvelopeWire(envelope), null, 2)}
 `,
     "utf8"
@@ -42305,13 +42842,13 @@ function buildResultIndexArtifact(result, extraIndexFields) {
     failure_reason_code: result.failureReasonCode,
     workspace_path: result.workspacePath,
     artifact_dir: artifactSubdir,
-    grading_path: path47.posix.join(artifactSubdir, "grading.json"),
-    timing_path: path47.posix.join(artifactSubdir, "timing.json"),
-    input_path: input ? path47.posix.join(artifactSubdir, "input.md") : void 0,
-    output_path: hasAnswer ? path47.posix.join(artifactSubdir, "outputs", "answer.md") : void 0,
-    answer_path: hasAnswer ? path47.posix.join(artifactSubdir, "outputs", "answer.md") : void 0,
-    transcript_path: hasTranscript ? path47.posix.join(artifactSubdir, "outputs", "transcript.jsonl") : void 0,
-    response_path: hasAnswer ? path47.posix.join(artifactSubdir, "outputs", "response.md") : void 0,
+    grading_path: path48.posix.join(artifactSubdir, "grading.json"),
+    timing_path: path48.posix.join(artifactSubdir, "timing.json"),
+    input_path: input ? path48.posix.join(artifactSubdir, "input.md") : void 0,
+    output_path: hasAnswer ? path48.posix.join(artifactSubdir, "outputs", "answer.md") : void 0,
+    answer_path: hasAnswer ? path48.posix.join(artifactSubdir, "outputs", "answer.md") : void 0,
+    transcript_path: hasTranscript ? path48.posix.join(artifactSubdir, "outputs", "transcript.jsonl") : void 0,
+    response_path: hasAnswer ? path48.posix.join(artifactSubdir, "outputs", "response.md") : void 0,
     ...extraIndexFields,
     metadata: toIndexMetadata(result.metadata)
   };
@@ -42351,7 +42888,7 @@ async function rewriteExistingIndexRecords(outputDir, replacements) {
   if (replacements.length === 0) {
     return;
   }
-  const indexPath = path47.join(outputDir, RESULT_INDEX_FILENAME);
+  const indexPath = path48.join(outputDir, RESULT_INDEX_FILENAME);
   const content = await readFile20(indexPath, "utf8").catch(() => void 0);
   if (content === void 0) {
     return;
@@ -42520,29 +43057,29 @@ async function writePerTestArtifacts(results, outputDir, options) {
     const grading = buildGradingArtifact(result);
     const timing = buildTimingArtifact([result]);
     const artifactSubdir = buildArtifactSubdir(result);
-    const testDir = path47.join(outputDir, artifactSubdir);
+    const testDir = path48.join(outputDir, artifactSubdir);
     await mkdir18(testDir, { recursive: true });
     await writeFile10(
-      path47.join(testDir, "grading.json"),
+      path48.join(testDir, "grading.json"),
       `${JSON.stringify(grading, null, 2)}
 `,
       "utf8"
     );
     await writeFile10(
-      path47.join(testDir, "timing.json"),
+      path48.join(testDir, "timing.json"),
       `${JSON.stringify(timing, null, 2)}
 `,
       "utf8"
     );
     const input = extractInput(result);
     if (input) {
-      await writeFile10(path47.join(testDir, "input.md"), input, "utf8");
+      await writeFile10(path48.join(testDir, "input.md"), input, "utf8");
     }
-    const outputsDir = path47.join(testDir, "outputs");
+    const outputsDir = path48.join(testDir, "outputs");
     await mkdir18(outputsDir, { recursive: true });
     if (result.output.length > 0) {
-      await writeFile10(path47.join(outputsDir, "answer.md"), result.output, "utf8");
-      await writeFile10(path47.join(outputsDir, "response.md"), result.output, "utf8");
+      await writeFile10(path48.join(outputsDir, "answer.md"), result.output, "utf8");
+      await writeFile10(path48.join(outputsDir, "response.md"), result.output, "utf8");
     }
     const envelope = await writeTraceEnvelopeSidecar({
       result,
@@ -42552,7 +43089,7 @@ async function writePerTestArtifacts(results, outputDir, options) {
       experiment: options?.experiment
     });
     if (hasTranscriptProjection(result, envelope)) {
-      await writeTranscriptJsonl(path47.join(outputsDir, "transcript.jsonl"), result, envelope);
+      await writeTranscriptJsonl(path48.join(outputsDir, "transcript.jsonl"), result, envelope);
     }
     const extraIndexFields = await collectAdditionalIndexFields(
       result,
@@ -42570,9 +43107,9 @@ async function writePerTestArtifacts(results, outputDir, options) {
 }
 async function writeArtifactsFromResults(results, outputDir, options) {
   const testArtifactDir = outputDir;
-  const timingPath = path47.join(outputDir, "timing.json");
-  const benchmarkPath = path47.join(outputDir, "benchmark.json");
-  const indexPath = path47.join(outputDir, RESULT_INDEX_FILENAME);
+  const timingPath = path48.join(outputDir, "timing.json");
+  const benchmarkPath = path48.join(outputDir, "benchmark.json");
+  const indexPath = path48.join(outputDir, RESULT_INDEX_FILENAME);
   await mkdir18(outputDir, { recursive: true });
   const indexRecords = [];
   const testByTestId = new Map((options?.sourceTests ?? []).map((test) => [test.id, test]));
@@ -42580,23 +43117,23 @@ async function writeArtifactsFromResults(results, outputDir, options) {
     const grading = buildGradingArtifact(result);
     const timing2 = buildTimingArtifact([result]);
     const artifactSubdir = buildArtifactSubdir(result);
-    const testDir = path47.join(outputDir, artifactSubdir);
-    const gradingPath = path47.join(testDir, "grading.json");
-    const perTestTimingPath = path47.join(testDir, "timing.json");
+    const testDir = path48.join(outputDir, artifactSubdir);
+    const gradingPath = path48.join(testDir, "grading.json");
+    const perTestTimingPath = path48.join(testDir, "timing.json");
     await mkdir18(testDir, { recursive: true });
     await writeFile10(gradingPath, `${JSON.stringify(grading, null, 2)}
 `, "utf8");
     await writeFile10(perTestTimingPath, `${JSON.stringify(timing2, null, 2)}
 `, "utf8");
     const input = extractInput(result);
-    const inputPath = input ? path47.join(testDir, "input.md") : void 0;
+    const inputPath = input ? path48.join(testDir, "input.md") : void 0;
     if (inputPath && input) {
       await writeFile10(inputPath, input, "utf8");
     }
-    const outputsDir = path47.join(testDir, "outputs");
+    const outputsDir = path48.join(testDir, "outputs");
     await mkdir18(outputsDir, { recursive: true });
-    const answerPath = result.output.length > 0 ? path47.join(outputsDir, "answer.md") : void 0;
-    const responsePath = result.output.length > 0 ? path47.join(outputsDir, "response.md") : void 0;
+    const answerPath = result.output.length > 0 ? path48.join(outputsDir, "answer.md") : void 0;
+    const responsePath = result.output.length > 0 ? path48.join(outputsDir, "response.md") : void 0;
     if (answerPath && responsePath) {
       await writeFile10(answerPath, result.output, "utf8");
       await writeFile10(responsePath, result.output, "utf8");
@@ -42608,7 +43145,7 @@ async function writeArtifactsFromResults(results, outputDir, options) {
       evalPath: resolveEnvelopeEvalPath(result, testByTestId, options?.evalFile),
       experiment: options?.experiment
     });
-    const transcriptPath = hasTranscriptProjection(result, envelope) ? path47.join(outputsDir, "transcript.jsonl") : void 0;
+    const transcriptPath = hasTranscriptProjection(result, envelope) ? path48.join(outputsDir, "transcript.jsonl") : void 0;
     if (transcriptPath) {
       await writeTranscriptJsonl(transcriptPath, result, envelope);
     }
@@ -42649,7 +43186,7 @@ async function writeArtifactsFromResults(results, outputDir, options) {
 `, "utf8");
   await writeJsonlFile(indexPath, indexRecords);
   await writeFile10(
-    path47.join(outputDir, "transcript.jsonl"),
+    path48.join(outputDir, "transcript.jsonl"),
     buildTranscriptMessageLines(results),
     "utf8"
   );
@@ -42700,7 +43237,7 @@ async function evaluate(config2) {
     cliNoCache: false,
     yamlCache: config2.cache === void 0 ? materialized.cache : void 0
   });
-  const cache = cacheEnabled ? new ResponseCache(materialized.cachePath ? path48.resolve(materialized.cachePath) : void 0) : void 0;
+  const cache = cacheEnabled ? new ResponseCache(materialized.cachePath ? path49.resolve(materialized.cachePath) : void 0) : void 0;
   const results = await runEvaluation({
     testFilePath,
     repoRoot,
@@ -42723,7 +43260,7 @@ async function evaluate(config2) {
   });
   const allResults = collectedResults.length > 0 ? collectedResults : [...results];
   const durationMs = Date.now() - startTime;
-  const outputDir = config2.outputDir ? path48.resolve(config2.outputDir) : void 0;
+  const outputDir = config2.outputDir ? path49.resolve(config2.outputDir) : void 0;
   const artifacts = outputDir ? await writeArtifactsFromResults(allResults, outputDir, {
     evalFile: config2.specFile ? testFilePath : "",
     experiment: config2.experiment,
@@ -42743,7 +43280,7 @@ async function evaluate(config2) {
 async function materializeEvalConfig(config2, options) {
   const baseDir = options?.baseDir ?? process.cwd();
   const repoRoot = options?.repoRoot ?? await findGitRoot(baseDir) ?? baseDir;
-  const testFilePath = config2.specFile ? path48.resolve(baseDir, config2.specFile) : path48.join(baseDir, "__programmatic__.yaml");
+  const testFilePath = config2.specFile ? path49.resolve(baseDir, config2.specFile) : path49.join(baseDir, "__programmatic__.yaml");
   const effectiveFilter = options?.filter ?? config2.filter;
   if (config2.specFile) {
     const suite = await loadTestSuite(testFilePath, repoRoot, {
@@ -42820,7 +43357,7 @@ function convertAssertions(entries) {
 }
 function buildInlineEvalTests(config2, options) {
   const suiteWorkspace = config2.beforeAll ? { hooks: { before_all: toBeforeAllHook(config2.beforeAll) } } : void 0;
-  const derivedSuiteName = path48.basename(options.testFilePath).replace(/\.eval\.[cm]?ts$/i, "").replace(/\.[cm]?ts$/i, "");
+  const derivedSuiteName = path49.basename(options.testFilePath).replace(/\.eval\.[cm]?ts$/i, "").replace(/\.[cm]?ts$/i, "");
   const suiteName = config2.metadata?.name ?? (derivedSuiteName || "eval");
   return (config2.tests ?? []).filter((test) => !options.filter || matchesFilter4(test.id, options.filter)).map((test) => {
     const isConversation = test.mode === "conversation" || test.turns && test.turns.length > 0;
@@ -42916,10 +43453,10 @@ function computeSummary(results, durationMs, threshold = DEFAULT_THRESHOLD) {
 var TARGET_FILE_CANDIDATES = [".agentv/targets.yaml", ".agentv/targets.yml"];
 async function discoverDefaultTarget(repoRoot) {
   const cwd = process.cwd();
-  const chain = buildDirectoryChain(path48.join(cwd, "_placeholder"), repoRoot);
+  const chain = buildDirectoryChain(path49.join(cwd, "_placeholder"), repoRoot);
   for (const dir of chain) {
     for (const candidate of TARGET_FILE_CANDIDATES) {
-      const targetsPath = path48.join(dir, candidate);
+      const targetsPath = path49.join(dir, candidate);
       if (!existsSync7(targetsPath)) continue;
       try {
         const definitions = await readTargetDefinitions(targetsPath);
@@ -42936,7 +43473,7 @@ async function loadEnvHierarchy(repoRoot, startPath) {
   const chain = buildDirectoryChain(startPath, repoRoot);
   const envFiles = [];
   for (const dir of chain) {
-    const envPath = path48.join(dir, ".env");
+    const envPath = path49.join(dir, ".env");
     if (existsSync7(envPath)) envFiles.push(envPath);
   }
   for (let i = 0; i < envFiles.length; i++) {
@@ -42962,7 +43499,7 @@ async function loadEnvHierarchy(repoRoot, startPath) {
 }
 var EXPORT_NAMES = ["default", "config", "evalConfig"];
 async function loadTsEvalFile(filePath) {
-  const absolutePath = path49.resolve(filePath);
+  const absolutePath = path50.resolve(filePath);
   const moduleUrl = pathToFileURL2(absolutePath).href;
   const module = await import(moduleUrl);
   let config2;
@@ -42984,7 +43521,7 @@ async function loadTsEvalSuite(filePath, repoRoot, options) {
   const { config: config2, filePath: absolutePath } = await loadTsEvalFile(filePath);
   const materialized = await materializeEvalConfig(config2, {
     repoRoot,
-    baseDir: path49.dirname(absolutePath),
+    baseDir: path50.dirname(absolutePath),
     filter: options?.filter,
     category: options?.category
   });
@@ -43046,6 +43583,7 @@ export {
   buildDirectoryChain,
   buildSearchRoots,
   resolveFileReference,
+  AGENT_PROVIDER_KINDS,
   KNOWN_PROVIDERS,
   PROVIDER_ALIASES,
   extractLastAssistantContent,
@@ -43230,6 +43768,7 @@ export {
   createTempWorkspace,
   cleanupWorkspace,
   cleanupEvalWorkspaces,
+  executeWorkspaceScript,
   resolveRepoCloneUrl,
   normalizeRepoIdentity,
   computeWorkspaceFingerprint,
@@ -43246,7 +43785,9 @@ export {
   discoverProjects,
   RepoManager,
   resolveWorkspaceTemplate,
-  executeWorkspaceScript,
+  releaseSharedWorkspaceSetup,
+  prepareSharedWorkspaceSetup,
+  prepareEvalCaseWorkspace,
   isAgentSkillsFormat,
   parseAgentSkillsEvals,
   DEFAULT_EVAL_PATTERNS,
@@ -43271,6 +43812,7 @@ export {
   loadEvalCases,
   loadTestById,
   loadEvalCaseById,
+  gradePreparedEvalCase,
   runEvaluation,
   runEvalCase,
   toTranscriptJsonLines,
@@ -43300,4 +43842,4 @@ export {
   loadTsEvalFile,
   loadTsEvalSuite
 };
-//# sourceMappingURL=chunk-BLXYBUU4.js.map
+//# sourceMappingURL=chunk-ENHX2CCS.js.map