@agentv/core 2.14.3 → 2.16.0

package/dist/index.cjs

@@ -1244,12 +1244,12 @@ function serializeAttributeValue(value) {
   if (Array.isArray(value)) return { arrayValue: { values: value.map(serializeAttributeValue) } };
   return { stringValue: String(value) };
 }
-var import_promises31, import_node_path44, OtlpJsonFileExporter;
+var import_promises31, import_node_path45, OtlpJsonFileExporter;
 var init_otlp_json_file_exporter = __esm({
   "src/observability/otlp-json-file-exporter.ts"() {
     "use strict";
     import_promises31 = require("fs/promises");
-    import_node_path44 = require("path");
+    import_node_path45 = require("path");
     OtlpJsonFileExporter = class {
       // biome-ignore lint/suspicious/noExplicitAny: serialized span data
       spans = [];
@@ -1288,7 +1288,7 @@ var init_otlp_json_file_exporter = __esm({
       }
       async flush() {
         if (this.spans.length === 0) return;
-        await (0, import_promises31.mkdir)((0, import_node_path44.dirname)(this.filePath), { recursive: true });
+        await (0, import_promises31.mkdir)((0, import_node_path45.dirname)(this.filePath), { recursive: true });
         const otlpJson = {
           resourceSpans: [
             {
@@ -1319,13 +1319,13 @@ function hrTimeDiffMs(start, end) {
   const diffNano = end[1] - start[1];
   return Math.round(diffSec * 1e3 + diffNano / 1e6);
 }
-var import_node_fs13, import_promises32, import_node_path45, SimpleTraceFileExporter;
+var import_node_fs13, import_promises32, import_node_path46, SimpleTraceFileExporter;
 var init_simple_trace_file_exporter = __esm({
   "src/observability/simple-trace-file-exporter.ts"() {
     "use strict";
     import_node_fs13 = require("fs");
     import_promises32 = require("fs/promises");
-    import_node_path45 = require("path");
+    import_node_path46 = require("path");
     SimpleTraceFileExporter = class {
       stream = null;
       filePath;
@@ -1338,7 +1338,7 @@ var init_simple_trace_file_exporter = __esm({
       async ensureStream() {
         if (!this.streamReady) {
           this.streamReady = (async () => {
-            await (0, import_promises32.mkdir)((0, import_node_path45.dirname)(this.filePath), { recursive: true });
+            await (0, import_promises32.mkdir)((0, import_node_path46.dirname)(this.filePath), { recursive: true });
             this.stream = (0, import_node_fs13.createWriteStream)(this.filePath, { flags: "w" });
             return this.stream;
           })();
@@ -1457,6 +1457,7 @@ __export(index_exports, {
   TokenUsageEvaluator: () => TokenUsageEvaluator,
   ToolTrajectoryEvaluator: () => ToolTrajectoryEvaluator,
   WorkspaceCreationError: () => WorkspaceCreationError,
+  WorkspacePoolManager: () => WorkspacePoolManager,
   assembleLlmJudgePrompt: () => assembleLlmJudgePrompt,
   avgToolDurationMs: () => avgToolDurationMs,
   buildDirectoryChain: () => buildDirectoryChain2,
@@ -1471,6 +1472,7 @@ __export(index_exports, {
   cleanupEvalWorkspaces: () => cleanupEvalWorkspaces,
   cleanupWorkspace: () => cleanupWorkspace,
   computeTraceSummary: () => computeTraceSummary,
+  computeWorkspaceFingerprint: () => computeWorkspaceFingerprint,
   consumeClaudeLogEntries: () => consumeClaudeLogEntries,
   consumeCodexLogEntries: () => consumeCodexLogEntries,
   consumeCopilotCliLogEntries: () => consumeCopilotCliLogEntries,
@@ -1503,11 +1505,11 @@ __export(index_exports, {
   freeformEvaluationSchema: () => freeformEvaluationSchema,
   generateRubrics: () => generateRubrics,
   getAgentvHome: () => getAgentvHome,
-  getGitCacheRoot: () => getGitCacheRoot,
   getHitCount: () => getHitCount,
   getSubagentsRoot: () => getSubagentsRoot,
   getTraceStateRoot: () => getTraceStateRoot,
   getWorkspacePath: () => getWorkspacePath,
+  getWorkspacePoolRoot: () => getWorkspacePoolRoot,
   getWorkspacesRoot: () => getWorkspacesRoot,
   initializeBaseline: () => initializeBaseline,
   isEvaluatorKind: () => isEvaluatorKind,
@@ -2236,6 +2238,17 @@ function parseExecutionDefaults(raw, configPath) {
   } else if (otelFile !== void 0) {
     logWarning(`Invalid execution.otel_file in ${configPath}, expected non-empty string`);
   }
+  if (typeof obj.pool_workspaces === "boolean") {
+    result.pool_workspaces = obj.pool_workspaces;
+  } else if (obj.pool_workspaces !== void 0) {
+    logWarning(`Invalid execution.pool_workspaces in ${configPath}, expected boolean`);
+  }
+  const poolSlots = obj.pool_slots;
+  if (typeof poolSlots === "number" && Number.isInteger(poolSlots) && poolSlots >= 1 && poolSlots <= 50) {
+    result.pool_slots = poolSlots;
+  } else if (poolSlots !== void 0) {
+    logWarning(`Invalid execution.pool_slots in ${configPath}, expected integer 1-50`);
+  }
   return Object.keys(result).length > 0 ? result : void 0;
 }
 function logWarning(message) {
@@ -3677,6 +3690,7 @@ async function processMessages(options) {
     repoRootPath,
     guidelinePatterns,
     guidelinePaths,
+    treatFileSegmentsAsGuidelines,
     textParts,
     messageType,
     verbose
@@ -3724,16 +3738,20 @@ async function processMessages(options) {
         }
         try {
           const fileContent = (await (0, import_promises5.readFile)(resolvedPath, "utf8")).replace(/\r\n/g, "\n");
-          if (messageType === "input" && guidelinePatterns && guidelinePaths) {
-            const relativeToRepo = import_node_path5.default.relative(repoRootPath, resolvedPath);
-            if (isGuidelineFile(relativeToRepo, guidelinePatterns)) {
-              guidelinePaths.push(import_node_path5.default.resolve(resolvedPath));
-              if (verbose) {
-                console.log(`  [Guideline] Found: ${displayPath}`);
-                console.log(`    Resolved to: ${resolvedPath}`);
-              }
-              continue;
+          const classifyAsGuideline = shouldTreatAsGuideline({
+            messageType,
+            resolvedPath,
+            repoRootPath,
+            guidelinePatterns,
+            treatFileSegmentsAsGuidelines
+          });
+          if (classifyAsGuideline && guidelinePaths) {
+            guidelinePaths.push(import_node_path5.default.resolve(resolvedPath));
+            if (verbose) {
+              console.log(`  [Guideline] Found: ${displayPath}`);
+              console.log(`    Resolved to: ${resolvedPath}`);
             }
+            continue;
           }
           segments.push({
             type: "file",
@@ -3762,6 +3780,26 @@ async function processMessages(options) {
   }
   return segments;
 }
+function shouldTreatAsGuideline(options) {
+  const {
+    messageType,
+    resolvedPath,
+    repoRootPath,
+    guidelinePatterns,
+    treatFileSegmentsAsGuidelines
+  } = options;
+  if (messageType !== "input") {
+    return false;
+  }
+  if (treatFileSegmentsAsGuidelines) {
+    return true;
+  }
+  if (!guidelinePatterns || guidelinePatterns.length === 0) {
+    return false;
+  }
+  const relativeToRepo = import_node_path5.default.relative(repoRootPath, resolvedPath);
+  return isGuidelineFile(relativeToRepo, guidelinePatterns);
+}
 function asString3(value) {
   return typeof value === "string" ? value : void 0;
 }
@@ -4100,6 +4138,8 @@ async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
         for (const guidelinePath of testCase.guideline_paths) {
           console.log(`    - ${guidelinePath}`);
         }
+      } else if (!guidelinePatterns || guidelinePatterns.length === 0) {
+        console.log("  No guidelines found (guideline_patterns not configured)");
       } else {
         console.log("  No guidelines found");
       }
@@ -4469,7 +4509,7 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
   } else {
     throw new Error(`Invalid test file format: ${evalFilePath} - missing 'tests' field`);
   }
-  const suiteWorkspace = parseWorkspaceConfig(suite.workspace, evalFileDir);
+  const suiteWorkspace = await resolveWorkspaceConfig(suite.workspace, evalFileDir);
   const suiteInputMessages = expandInputShorthand(suite.input);
   const rawGlobalExecution = isJsonObject(suite.execution) ? suite.execution : void 0;
   const _globalTarget = asString6(rawGlobalExecution?.target) ?? asString6(suite.target);
@@ -4505,12 +4545,24 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
     }
     const caseExecution = isJsonObject(evalcase.execution) ? evalcase.execution : void 0;
     const skipDefaults = caseExecution?.skip_defaults === true;
-    const inputMessages = suiteInputMessages && !skipDefaults ? [...suiteInputMessages, ...testInputMessages] : testInputMessages;
+    const effectiveSuiteInputMessages = suiteInputMessages && !skipDefaults ? suiteInputMessages : void 0;
+    const inputMessages = effectiveSuiteInputMessages ? [...effectiveSuiteInputMessages, ...testInputMessages] : testInputMessages;
     const hasExpectedMessages = expectedMessages.length > 0;
     const guidelinePaths = [];
     const inputTextParts = [];
-    const inputSegments = await processMessages({
-      messages: inputMessages,
+    const suiteInputSegments = effectiveSuiteInputMessages ? await processMessages({
+      messages: effectiveSuiteInputMessages,
+      searchRoots,
+      repoRootPath,
+      guidelinePatterns,
+      guidelinePaths,
+      treatFileSegmentsAsGuidelines: true,
+      textParts: inputTextParts,
+      messageType: "input",
+      verbose
+    }) : [];
+    const testInputSegments = await processMessages({
+      messages: testInputMessages,
       searchRoots,
       repoRootPath,
       guidelinePatterns,
@@ -4519,6 +4571,7 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
       messageType: "input",
       verbose
     });
+    const inputSegments = [...suiteInputSegments, ...testInputSegments];
     const outputSegments = hasExpectedMessages ? await processExpectedMessages({
       messages: expectedMessages,
       searchRoots,
@@ -4566,7 +4619,7 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
       ...guidelinePaths.map((guidelinePath) => import_node_path8.default.resolve(guidelinePath)),
       ...userFilePaths
     ];
-    const caseWorkspace = parseWorkspaceConfig(evalcase.workspace, evalFileDir);
+    const caseWorkspace = await resolveWorkspaceConfig(evalcase.workspace, evalFileDir);
     const mergedWorkspace = mergeWorkspaceConfigs(suiteWorkspace, caseWorkspace);
     const metadata = isJsonObject(evalcase.metadata) ? evalcase.metadata : void 0;
     const caseTargets = extractTargetsFromTestCase(evalcase);
@@ -4597,6 +4650,8 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
         for (const guidelinePath of testCase.guideline_paths) {
           console.log(`    - ${guidelinePath}`);
         }
+      } else if (!guidelinePatterns || guidelinePatterns.length === 0) {
+        console.log("  No guidelines found (guideline_patterns not configured)");
       } else {
         console.log("  No guidelines found");
       }
@@ -4685,16 +4740,57 @@ function parseRepoConfig(raw) {
     ...clone !== void 0 && { clone }
   };
 }
-function parseResetConfig(raw) {
+function parseWorkspaceHookConfig(raw, evalFileDir) {
   if (!isJsonObject(raw)) return void 0;
+  const script = parseWorkspaceScriptConfig(raw, evalFileDir);
   const obj = raw;
-  const strategy = obj.strategy === "none" || obj.strategy === "hard" || obj.strategy === "recreate" ? obj.strategy : void 0;
-  const afterEach = typeof obj.after_each === "boolean" ? obj.after_each : void 0;
-  if (!strategy && afterEach === void 0) return void 0;
+  const reset = obj.reset === "none" || obj.reset === "fast" || obj.reset === "strict" ? obj.reset : void 0;
+  const clean = obj.clean === "always" || obj.clean === "on_success" || obj.clean === "on_failure" || obj.clean === "never" ? obj.clean : void 0;
+  if (!script && !reset && !clean) return void 0;
   return {
-    ...strategy !== void 0 && { strategy },
-    ...afterEach !== void 0 && { after_each: afterEach }
+    ...script ?? {},
+    ...reset !== void 0 && { reset },
+    ...clean !== void 0 && { clean }
+  };
+}
+function parseWorkspaceHooksConfig(raw, evalFileDir) {
+  if (!isJsonObject(raw)) return void 0;
+  const obj = raw;
+  const beforeAllTests = parseWorkspaceHookConfig(obj.before_all_tests, evalFileDir);
+  const beforeEachTest = parseWorkspaceHookConfig(obj.before_each_test, evalFileDir);
+  const afterEachTest = parseWorkspaceHookConfig(obj.after_each_test, evalFileDir);
+  const afterAllTests = parseWorkspaceHookConfig(obj.after_all_tests, evalFileDir);
+  const onReuse = parseWorkspaceHookConfig(obj.on_reuse, evalFileDir);
+  const onFinish = parseWorkspaceHookConfig(obj.on_finish, evalFileDir);
+  const hooks = {
+    ...beforeAllTests !== void 0 && { before_all_tests: beforeAllTests },
+    ...beforeEachTest !== void 0 && { before_each_test: beforeEachTest },
+    ...afterEachTest !== void 0 && { after_each_test: afterEachTest },
+    ...afterAllTests !== void 0 && { after_all_tests: afterAllTests },
+    ...onReuse !== void 0 && { on_reuse: onReuse },
+    ...onFinish !== void 0 && { on_finish: onFinish }
   };
+  return Object.keys(hooks).length > 0 ? hooks : void 0;
+}
+async function resolveWorkspaceConfig(raw, evalFileDir) {
+  if (typeof raw === "string") {
+    const workspaceFilePath = import_node_path8.default.resolve(evalFileDir, raw);
+    let content;
+    try {
+      content = await (0, import_promises8.readFile)(workspaceFilePath, "utf8");
+    } catch {
+      throw new Error(`Workspace file not found: ${raw} (resolved to ${workspaceFilePath})`);
+    }
+    const parsed = (0, import_yaml4.parse)(content);
+    if (!isJsonObject(parsed)) {
+      throw new Error(
+        `Invalid workspace file format: ${workspaceFilePath} (expected a YAML object)`
+      );
+    }
+    const workspaceFileDir = import_node_path8.default.dirname(workspaceFilePath);
+    return parseWorkspaceConfig(parsed, workspaceFileDir);
+  }
+  return parseWorkspaceConfig(raw, evalFileDir);
 }
 function parseWorkspaceConfig(raw, evalFileDir) {
   if (!isJsonObject(raw)) return void 0;
@@ -4705,37 +4801,56 @@ function parseWorkspaceConfig(raw, evalFileDir) {
   }
   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;
-  const reset = parseResetConfig(obj.reset);
-  const beforeAll = parseWorkspaceScriptConfig(obj.before_all, evalFileDir);
-  const afterAll = parseWorkspaceScriptConfig(obj.after_all, evalFileDir);
-  const beforeEach = parseWorkspaceScriptConfig(obj.before_each, evalFileDir);
-  const afterEach = parseWorkspaceScriptConfig(obj.after_each, evalFileDir);
-  if (!template && !isolation && !repos && !reset && !beforeAll && !afterAll && !beforeEach && !afterEach)
+  const hooks = parseWorkspaceHooksConfig(obj.hooks, evalFileDir);
+  const mode = obj.mode === "pooled" || obj.mode === "ephemeral" || obj.mode === "static" ? obj.mode : void 0;
+  const staticPath = typeof obj.static_path === "string" ? obj.static_path : void 0;
+  const pool = typeof obj.pool === "boolean" ? obj.pool : void 0;
+  if (!template && !isolation && !repos && !hooks && !mode && !staticPath && pool === void 0)
     return void 0;
   return {
     ...template !== void 0 && { template },
     ...isolation !== void 0 && { isolation },
     ...repos !== void 0 && { repos },
-    ...reset !== void 0 && { reset },
-    ...beforeAll !== void 0 && { before_all: beforeAll },
-    ...afterAll !== void 0 && { after_all: afterAll },
-    ...beforeEach !== void 0 && { before_each: beforeEach },
-    ...afterEach !== void 0 && { after_each: afterEach }
+    ...hooks !== void 0 && { hooks },
+    ...mode !== void 0 && { mode },
+    ...staticPath !== void 0 && { static_path: staticPath },
+    ...pool !== void 0 && { pool }
   };
 }
 function mergeWorkspaceConfigs(suiteLevel, caseLevel) {
   if (!suiteLevel && !caseLevel) return void 0;
   if (!suiteLevel) return caseLevel;
   if (!caseLevel) return suiteLevel;
+  const mergeHook = (suiteHook, caseHook) => {
+    if (!suiteHook && !caseHook) return void 0;
+    return {
+      ...suiteHook ?? {},
+      ...caseHook ?? {}
+    };
+  };
+  const mergedHooks = {
+    before_all_tests: mergeHook(
+      suiteLevel.hooks?.before_all_tests,
+      caseLevel.hooks?.before_all_tests
+    ),
+    before_each_test: mergeHook(
+      suiteLevel.hooks?.before_each_test,
+      caseLevel.hooks?.before_each_test
+    ),
+    after_each_test: mergeHook(suiteLevel.hooks?.after_each_test, caseLevel.hooks?.after_each_test),
+    after_all_tests: mergeHook(suiteLevel.hooks?.after_all_tests, caseLevel.hooks?.after_all_tests),
+    on_reuse: mergeHook(suiteLevel.hooks?.on_reuse, caseLevel.hooks?.on_reuse),
+    on_finish: mergeHook(suiteLevel.hooks?.on_finish, caseLevel.hooks?.on_finish)
+  };
+  const hasHooks = Object.values(mergedHooks).some((hook) => hook !== void 0);
   return {
     template: caseLevel.template ?? suiteLevel.template,
     isolation: caseLevel.isolation ?? suiteLevel.isolation,
     repos: caseLevel.repos ?? suiteLevel.repos,
-    reset: caseLevel.reset ?? suiteLevel.reset,
-    before_all: caseLevel.before_all ?? suiteLevel.before_all,
-    after_all: caseLevel.after_all ?? suiteLevel.after_all,
-    before_each: caseLevel.before_each ?? suiteLevel.before_each,
-    after_each: caseLevel.after_each ?? suiteLevel.after_each
+    ...hasHooks && { hooks: mergedHooks },
+    mode: caseLevel.mode ?? suiteLevel.mode,
+    static_path: caseLevel.static_path ?? suiteLevel.static_path,
+    pool: caseLevel.pool ?? suiteLevel.pool
   };
 }
 function asString6(value) {
@@ -9493,8 +9608,8 @@ function resolveCliConfig(target, env, evalFilePath) {
   const parseResult = CliTargetInputSchema.safeParse(target, { errorMap: cliErrorMap });
   if (!parseResult.success) {
     const firstError = parseResult.error.errors[0];
-    const path43 = firstError?.path.join(".") || "";
-    const prefix = path43 ? `${target.name} ${path43}: ` : `${target.name}: `;
+    const path44 = firstError?.path.join(".") || "";
+    const prefix = path44 ? `${target.name} ${path44}: ` : `${target.name}: `;
     throw new Error(`${prefix}${firstError?.message}`);
   }
   const normalized = normalizeCliTargetInput(parseResult.data, env, evalFilePath);
@@ -10001,15 +10116,15 @@ function getAgentvHome() {
 function getWorkspacesRoot() {
   return import_node_path23.default.join(getAgentvHome(), "workspaces");
 }
-function getGitCacheRoot() {
-  return import_node_path23.default.join(getAgentvHome(), "git-cache");
-}
 function getSubagentsRoot() {
   return import_node_path23.default.join(getAgentvHome(), "subagents");
 }
 function getTraceStateRoot() {
   return import_node_path23.default.join(getAgentvHome(), "trace-state");
 }
+function getWorkspacePoolRoot() {
+  return import_node_path23.default.join(getAgentvHome(), "workspace-pool");
+}
 
 // src/evaluation/providers/vscode/dispatch/constants.ts
 var DEFAULT_LOCK_NAME = "subagent.lock";
@@ -10832,8 +10947,6 @@ var AGENTV_REQUEST_TEMPLATE = `[[ ## task ## ]]
 
 **IMPORTANT**: Follow these exact steps:
 1. Create and write your complete response to: {{responseFileTmp}}
-    - All intended file outputs/changes MUST be written in your response file.
-    - For each intended file, include the repo name, relative path and unified git diff following the convention \`diff --git ...\`.
 2. When completely finished, run these PowerShell commands to signal completion:
 \`\`\`
 Move-Item -LiteralPath '{{responseFileTmp}}' -Destination '{{responseFileFinal}}'
@@ -10850,8 +10963,6 @@ var AGENTV_BATCH_REQUEST_TEMPLATE = `[[ ## task ## ]]
 
 **IMPORTANT**: Follow these exact steps:
 1. Create and write your complete response to: {{responseFileTmp}}
-    - All intended file outputs/changes MUST be written in your response file.
-    - For each intended file, include the repo name, relative path and unified git diff following the convention \`diff --git ...\`.
 2. When completely finished and the response is stable, rename it to: {{responseFileFinal}}
 3. Do not unlock the workspace from this request; batch orchestration will handle unlocking after all responses are ready.
 `;
@@ -11464,15 +11575,15 @@ async function execFileWithStdinNode(argv, stdinPayload, options) {
   });
 }
 async function execShellWithStdin(command, stdinPayload, options = {}) {
-  const { mkdir: mkdir16, readFile: readFile13, rm: rm6, writeFile: writeFile9 } = await import("fs/promises");
+  const { mkdir: mkdir16, readFile: readFile14, rm: rm6, writeFile: writeFile9 } = await import("fs/promises");
   const { tmpdir: tmpdir3 } = await import("os");
-  const path43 = await import("path");
+  const path44 = await import("path");
   const { randomUUID: randomUUID8 } = await import("crypto");
-  const dir = path43.join(tmpdir3(), `agentv-exec-${randomUUID8()}`);
+  const dir = path44.join(tmpdir3(), `agentv-exec-${randomUUID8()}`);
   await mkdir16(dir, { recursive: true });
-  const stdinPath = path43.join(dir, "stdin.txt");
-  const stdoutPath = path43.join(dir, "stdout.txt");
-  const stderrPath = path43.join(dir, "stderr.txt");
+  const stdinPath = path44.join(dir, "stdin.txt");
+  const stdoutPath = path44.join(dir, "stdout.txt");
+  const stderrPath = path44.join(dir, "stderr.txt");
   await writeFile9(stdinPath, stdinPayload, "utf8");
   const wrappedCommand = process.platform === "win32" ? `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}` : `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}`;
   const { spawn: spawn4 } = await import("child_process");
@@ -11502,8 +11613,8 @@ async function execShellWithStdin(command, stdinPayload, options = {}) {
         resolve(code ?? 0);
       });
     });
-    const stdout = (await readFile13(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
-    const stderr = (await readFile13(stderrPath, "utf8")).replace(/\r\n/g, "\n");
+    const stdout = (await readFile14(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
+    const stderr = (await readFile14(stderrPath, "utf8")).replace(/\r\n/g, "\n");
     return { stdout, stderr, exitCode };
   } finally {
     await rm6(dir, { recursive: true, force: true });
@@ -11824,7 +11935,7 @@ var CodeEvaluator = class {
       outputPath,
       guidelineFiles: context2.evalCase.guideline_paths,
       inputFiles: context2.evalCase.file_paths.filter(
-        (path43) => !context2.evalCase.guideline_paths.includes(path43)
+        (path44) => !context2.evalCase.guideline_paths.includes(path44)
       ),
       input: context2.evalCase.input,
       trace: context2.trace ?? null,
@@ -12103,6 +12214,8 @@ ${context2.fileChanges}`;
       };
     } catch (e) {
       const message = e instanceof Error ? e.message : String(e);
+      const evalName = context2.evaluator?.name ?? "llm-judge";
+      console.warn(`\u26A0 LLM judge "${evalName}" failed after 3 attempts (${message}) \u2014 skipped`);
       return {
         score: 0,
         verdict: "skip",
@@ -12131,24 +12244,39 @@ ${context2.fileChanges}`;
       systemPrompt,
       target: judgeProvider.targetName
     };
-    const { data, tokenUsage } = await this.runWithRetry({
-      context: context2,
-      judgeProvider,
-      systemPrompt,
-      userPrompt: prompt,
-      schema: rubricEvaluationSchema
-    });
-    const { score, verdict, hits, misses } = calculateRubricScore(data, rubrics);
-    return {
-      score,
-      verdict,
-      hits,
-      misses,
-      expectedAspectCount: rubrics.length,
-      reasoning: data.overall_reasoning,
-      evaluatorRawRequest,
-      tokenUsage
-    };
+    try {
+      const { data, tokenUsage } = await this.runWithRetry({
+        context: context2,
+        judgeProvider,
+        systemPrompt,
+        userPrompt: prompt,
+        schema: rubricEvaluationSchema
+      });
+      const { score, verdict, hits, misses } = calculateRubricScore(data, rubrics);
+      return {
+        score,
+        verdict,
+        hits,
+        misses,
+        expectedAspectCount: rubrics.length,
+        reasoning: data.overall_reasoning,
+        evaluatorRawRequest,
+        tokenUsage
+      };
+    } catch (e) {
+      const message = e instanceof Error ? e.message : String(e);
+      const evalName = context2.evaluator?.name ?? "llm-judge";
+      console.warn(`\u26A0 LLM judge "${evalName}" failed after 3 attempts (${message}) \u2014 skipped`);
+      return {
+        score: 0,
+        verdict: "skip",
+        hits: [],
+        misses: [`Judge parse failure after 3 attempts: ${message}`],
+        expectedAspectCount: rubrics.length,
+        reasoning: `Judge parse failure after 3 attempts: ${message}`,
+        evaluatorRawRequest
+      };
+    }
   }
   /**
    * Evaluate using score-range rubrics (analytic rubric scoring).
@@ -12162,25 +12290,40 @@ ${context2.fileChanges}`;
       systemPrompt,
       target: judgeProvider.targetName
     };
-    const { data, tokenUsage } = await this.runWithRetry({
-      context: context2,
-      judgeProvider,
-      systemPrompt,
-      userPrompt: prompt,
-      schema: scoreRangeEvaluationSchema
-    });
-    const { score, verdict, hits, misses, details } = calculateScoreRangeResult(data, rubrics);
-    return {
-      score,
-      verdict,
-      hits,
-      misses,
-      expectedAspectCount: rubrics.length,
-      reasoning: data.overall_reasoning,
-      evaluatorRawRequest,
-      details,
-      tokenUsage
-    };
+    try {
+      const { data, tokenUsage } = await this.runWithRetry({
+        context: context2,
+        judgeProvider,
+        systemPrompt,
+        userPrompt: prompt,
+        schema: scoreRangeEvaluationSchema
+      });
+      const { score, verdict, hits, misses, details } = calculateScoreRangeResult(data, rubrics);
+      return {
+        score,
+        verdict,
+        hits,
+        misses,
+        expectedAspectCount: rubrics.length,
+        reasoning: data.overall_reasoning,
+        evaluatorRawRequest,
+        details,
+        tokenUsage
+      };
+    } catch (e) {
+      const message = e instanceof Error ? e.message : String(e);
+      const evalName = context2.evaluator?.name ?? "llm-judge";
+      console.warn(`\u26A0 LLM judge "${evalName}" failed after 3 attempts (${message}) \u2014 skipped`);
+      return {
+        score: 0,
+        verdict: "skip",
+        hits: [],
+        misses: [`Judge parse failure after 3 attempts: ${message}`],
+        expectedAspectCount: rubrics.length,
+        reasoning: `Judge parse failure after 3 attempts: ${message}`,
+        evaluatorRawRequest
+      };
+    }
   }
   /**
    * Build prompt for score-range rubric evaluation.
@@ -12466,19 +12609,13 @@ var CompositeEvaluator = class {
   runWeightedAverage(results, weights) {
     let totalWeight = 0;
     let weightedSum = 0;
+    let evaluatedCount = 0;
     const allHits = [];
     const allMisses = [];
     const reasoningParts = [];
     const scores = [];
     for (const member of results) {
       const weight = weights?.[member.id] ?? 1;
-      totalWeight += weight;
-      weightedSum += member.result.score * weight;
-      allHits.push(...member.result.hits.map((h) => `[${member.id}] ${h}`));
-      allMisses.push(...member.result.misses.map((m) => `[${member.id}] ${m}`));
-      if (member.result.reasoning) {
-        reasoningParts.push(`${member.id}: ${member.result.reasoning}`);
-      }
       scores.push({
         name: member.id,
         type: member.type,
@@ -12493,6 +12630,32 @@ var CompositeEvaluator = class {
         details: member.result.details,
         tokenUsage: member.result.tokenUsage
       });
+      if (member.result.verdict === "skip") {
+        continue;
+      }
+      evaluatedCount++;
+      totalWeight += weight;
+      weightedSum += member.result.score * weight;
+      allHits.push(...member.result.hits.map((h) => `[${member.id}] ${h}`));
+      allMisses.push(...member.result.misses.map((m) => `[${member.id}] ${m}`));
+      if (member.result.reasoning) {
+        reasoningParts.push(`${member.id}: ${member.result.reasoning}`);
+      }
+    }
+    if (evaluatedCount === 0 && results.length > 0) {
+      return {
+        score: 0,
+        verdict: "skip",
+        hits: [],
+        misses: [],
+        expectedAspectCount: 1,
+        reasoning: "All evaluators skipped (infrastructure failure)",
+        evaluatorRawRequest: {
+          aggregator: "weighted_average",
+          ...weights ? { weights } : {}
+        },
+        scores
+      };
     }
     const finalScore = totalWeight > 0 ? weightedSum / totalWeight : 0;
     return {
@@ -12516,19 +12679,8 @@ var CompositeEvaluator = class {
     const reasoningParts = [];
     let passingCount = 0;
     let borderlineCount = 0;
+    let evaluatedCount = 0;
     for (const member of results) {
-      const isPassing = member.result.verdict === "pass" || member.result.verdict === "borderline";
-      if (isPassing) {
-        passingCount++;
-        if (member.result.verdict === "borderline") {
-          borderlineCount++;
-        }
-      }
-      allHits.push(...member.result.hits.map((h) => `[${member.id}] ${h}`));
-      allMisses.push(...member.result.misses.map((m) => `[${member.id}] ${m}`));
-      if (member.result.reasoning) {
-        reasoningParts.push(`${member.id}: ${member.result.reasoning}`);
-      }
       scores.push({
         name: member.id,
         type: member.type,
@@ -12542,8 +12694,39 @@ var CompositeEvaluator = class {
         details: member.result.details,
         tokenUsage: member.result.tokenUsage
       });
+      if (member.result.verdict === "skip") {
+        continue;
+      }
+      evaluatedCount++;
+      const isPassing = member.result.verdict === "pass" || member.result.verdict === "borderline";
+      if (isPassing) {
+        passingCount++;
+        if (member.result.verdict === "borderline") {
+          borderlineCount++;
+        }
+      }
+      allHits.push(...member.result.hits.map((h) => `[${member.id}] ${h}`));
+      allMisses.push(...member.result.misses.map((m) => `[${member.id}] ${m}`));
+      if (member.result.reasoning) {
+        reasoningParts.push(`${member.id}: ${member.result.reasoning}`);
+      }
+    }
+    if (evaluatedCount === 0 && results.length > 0) {
+      return {
+        score: 0,
+        verdict: "skip",
+        hits: [],
+        misses: [],
+        expectedAspectCount: 1,
+        reasoning: "All evaluators skipped (infrastructure failure)",
+        evaluatorRawRequest: {
+          aggregator: "threshold",
+          threshold
+        },
+        scores
+      };
     }
-    const totalCount = results.length;
+    const totalCount = evaluatedCount;
     const score = totalCount > 0 ? passingCount / totalCount : 0;
     const pass = score >= threshold;
     if (pass && borderlineCount > 0) {
@@ -13051,115 +13234,115 @@ var FieldAccuracyEvaluator = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path43, match, required = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path43);
-    const expectedValue = resolvePath(expectedData, path43);
+    const { path: path44, match, required = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path44);
+    const expectedValue = resolvePath(expectedData, path44);
     if (expectedValue === void 0) {
       return {
-        path: path43,
+        path: path44,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path43}: no expected value`
+        message: `${path44}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required) {
         return {
-          path: path43,
+          path: path44,
           score: 0,
           weight,
           hit: false,
-          message: `${path43} (required, missing)`
+          message: `${path44} (required, missing)`
         };
       }
       return {
-        path: path43,
+        path: path44,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path43}: optional field missing`
+        message: `${path44}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path43, candidateValue, expectedValue, weight);
+        return this.compareExact(path44, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path43,
+          path44,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path43, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path44, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path43,
+          path: path44,
           score: 0,
           weight,
           hit: false,
-          message: `${path43}: unknown match type "${match}"`
+          message: `${path44}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path43, candidateValue, expectedValue, weight) {
+  compareExact(path44, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path43,
+        path: path44,
         score: 1,
         weight,
         hit: true,
-        message: path43
+        message: path44
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path43,
+        path: path44,
         score: 0,
         weight,
         hit: false,
-        message: `${path43} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path44} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path43,
+      path: path44,
       score: 0,
       weight,
       hit: false,
-      message: `${path43} (value mismatch)`
+      message: `${path44} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path43, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path44, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber2(candidateValue);
     const expectedNum = toNumber2(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path43,
+        path: path44,
         score: 0,
         weight,
         hit: false,
-        message: `${path43} (non-numeric value)`
+        message: `${path44} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path43,
+        path: path44,
         score: 0,
         weight,
         hit: false,
-        message: `${path43} (invalid numeric value)`
+        message: `${path44} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -13172,61 +13355,61 @@ var FieldAccuracyEvaluator = class {
     }
     if (withinTolerance) {
       return {
-        path: path43,
+        path: path44,
         score: 1,
         weight,
         hit: true,
-        message: `${path43} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path44} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path43,
+      path: path44,
       score: 0,
       weight,
       hit: false,
-      message: `${path43} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path44} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path43, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path44, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path43,
+        path: path44,
         score: 0,
         weight,
         hit: false,
-        message: `${path43} (unparseable candidate date)`
+        message: `${path44} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path43,
+        path: path44,
         score: 0,
         weight,
         hit: false,
-        message: `${path43} (unparseable expected date)`
+        message: `${path44} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path43,
+        path: path44,
         score: 1,
         weight,
         hit: true,
-        message: path43
+        message: path44
       };
     }
     return {
-      path: path43,
+      path: path44,
       score: 0,
       weight,
       hit: false,
-      message: `${path43} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path44} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -13267,11 +13450,11 @@ var FieldAccuracyEvaluator = class {
     };
   }
 };
-function resolvePath(obj, path43) {
-  if (!path43 || !obj) {
+function resolvePath(obj, path44) {
+  if (!path44 || !obj) {
     return void 0;
   }
-  const parts = path43.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path44.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -14089,8 +14272,8 @@ var TokenUsageEvaluator = class {
 };
 
 // src/evaluation/evaluators/tool-trajectory.ts
-function getNestedValue(obj, path43) {
-  const parts = path43.split(".");
+function getNestedValue(obj, path44) {
+  const parts = path44.split(".");
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0 || typeof current !== "object") {
@@ -14653,7 +14836,7 @@ function runEqualsAssertion(output, value) {
 // src/evaluation/orchestrator.ts
 var import_node_crypto9 = require("crypto");
 var import_promises29 = require("fs/promises");
-var import_node_path41 = __toESM(require("path"), 1);
+var import_node_path42 = __toESM(require("path"), 1);
 var import_micromatch4 = __toESM(require("micromatch"), 1);
 
 // ../../node_modules/.bun/yocto-queue@1.2.2/node_modules/yocto-queue/index.js
@@ -15523,7 +15706,7 @@ async function cleanupEvalWorkspaces(evalRunId, workspaceRoot) {
   }
 }
 
-// src/evaluation/workspace/repo-manager.ts
+// src/evaluation/workspace/pool-manager.ts
 var import_node_child_process7 = require("child_process");
 var import_node_crypto8 = require("crypto");
 var import_node_fs11 = require("fs");
@@ -15531,8 +15714,6 @@ var import_promises27 = require("fs/promises");
 var import_node_path39 = __toESM(require("path"), 1);
 var import_node_util5 = require("util");
 var execFileAsync = (0, import_node_util5.promisify)(import_node_child_process7.execFile);
-var DEFAULT_TIMEOUT_MS2 = 3e5;
-var LOCK_TIMEOUT_MS = 6e4;
 function gitEnv() {
   const env = { ...process.env };
   for (const key of Object.keys(env)) {
@@ -15547,160 +15728,326 @@ function gitEnv() {
     GIT_SSH_COMMAND: "ssh -o BatchMode=yes"
   };
 }
-function cacheKey(source) {
-  const raw = source.type === "git" ? source.url.toLowerCase().replace(/\.git$/, "") : source.path;
-  return (0, import_node_crypto8.createHash)("sha256").update(raw).digest("hex");
-}
-function getSourceUrl(source) {
-  return source.type === "git" ? source.url : source.path;
-}
 async function git(args, opts) {
   const { stdout } = await execFileAsync("git", args, {
     cwd: opts?.cwd,
-    timeout: opts?.timeout ?? DEFAULT_TIMEOUT_MS2,
+    timeout: opts?.timeout ?? 3e5,
     env: gitEnv(),
     maxBuffer: 50 * 1024 * 1024
-    // 50MB
   });
   return stdout.trim();
 }
-async function acquireLock(lockPath) {
-  const start = Date.now();
-  while (Date.now() - start < LOCK_TIMEOUT_MS) {
-    try {
-      await (0, import_promises27.writeFile)(lockPath, String(process.pid), { flag: "wx" });
-      return;
-    } catch (err) {
-      if (err.code === "EEXIST") {
-        await new Promise((r) => setTimeout(r, 200));
+function normalizeRepoForFingerprint(repo) {
+  const source = repo.source.type === "git" ? { type: "git", url: repo.source.url.toLowerCase().replace(/\.git$/, "") } : { type: "local", path: repo.source.path };
+  const result = {
+    path: repo.path,
+    source,
+    ref: repo.checkout?.ref ?? "HEAD"
+  };
+  if (repo.clone?.depth !== void 0) {
+    result.depth = repo.clone.depth;
+  }
+  if (repo.clone?.filter !== void 0) {
+    result.filter = repo.clone.filter;
+  }
+  if (repo.clone?.sparse?.length) {
+    result.sparse = [...repo.clone.sparse].sort();
+  }
+  return result;
+}
+function computeWorkspaceFingerprint(templatePath, repos) {
+  const canonical = {
+    templatePath: templatePath ?? null,
+    repos: [...repos].sort((a, b) => a.path.localeCompare(b.path)).map(normalizeRepoForFingerprint)
+  };
+  return (0, import_node_crypto8.createHash)("sha256").update(JSON.stringify(canonical)).digest("hex");
+}
+async function copyDirectoryRecursive2(src, dest, skipDirs) {
+  await (0, import_promises27.mkdir)(dest, { recursive: true });
+  const entries = await (0, import_promises27.readdir)(src, { withFileTypes: true });
+  for (const entry of entries) {
+    const srcPath = import_node_path39.default.join(src, entry.name);
+    const destPath = import_node_path39.default.join(dest, entry.name);
+    if (entry.name === ".git") {
+      continue;
+    }
+    if (entry.isDirectory()) {
+      if (skipDirs?.has(entry.name)) {
         continue;
       }
-      throw err;
+      await copyDirectoryRecursive2(srcPath, destPath, skipDirs);
+    } else {
+      await (0, import_promises27.cp)(srcPath, destPath, { preserveTimestamps: true, force: true });
     }
   }
-  throw new Error(`Timed out waiting for lock: ${lockPath}`);
 }
-async function releaseLock(lockPath) {
-  try {
-    await (0, import_promises27.unlink)(lockPath);
-  } catch {
-  }
-}
-var RepoManager = class {
-  cacheDir;
-  verbose;
-  constructor(cacheDir, verbose = false) {
-    this.cacheDir = cacheDir ?? getGitCacheRoot();
-    this.verbose = verbose;
+var WorkspacePoolManager = class {
+  poolRoot;
+  constructor(poolRoot) {
+    this.poolRoot = poolRoot ?? getWorkspacePoolRoot();
   }
-  async runGit(args, opts) {
-    const startedAt = Date.now();
-    if (this.verbose) {
-      console.log(
-        `[repo] git start cwd=${opts?.cwd ?? process.cwd()} args=${args.join(" ")}`
+  /**
+   * Acquire a workspace slot from the pool.
+   *
+   * 1. Compute fingerprint from template + repos
+   * 2. Check drift (compare stored metadata.json fingerprint vs computed)
+   * 3. If drift: warn, remove all slots, rematerialize
+   * 4. Acquire a slot (try-lock slot-0, slot-1, ..., up to maxSlots)
+   * 5. If slot exists: reset repos, re-copy template files (skip repo directories)
+   * 6. If new slot: copy template, materialize all repos, write metadata.json
+   * 7. Return the slot (with path, index, isExisting)
+   */
+  async acquireWorkspace(options) {
+    const { templatePath, repos, maxSlots, repoManager, poolReset } = options;
+    const fingerprint = computeWorkspaceFingerprint(templatePath, repos);
+    const poolDir = import_node_path39.default.join(this.poolRoot, fingerprint);
+    await (0, import_promises27.mkdir)(poolDir, { recursive: true });
+    const drifted = await this.checkDrift(poolDir, fingerprint);
+    if (drifted) {
+      console.warn(
+        `[workspace-pool] Drift detected for fingerprint ${fingerprint.slice(0, 12)}... Removing stale slots.`
       );
+      await this.removeAllSlots(poolDir);
     }
-    try {
-      const output = await git(args, opts);
-      if (this.verbose) {
-        console.log(
-          `[repo] git ok durationMs=${Date.now() - startedAt} args=${args.join(" ")}`
-        );
+    for (let i = 0; i < maxSlots; i++) {
+      const slotPath = import_node_path39.default.join(poolDir, `slot-${i}`);
+      const lockPath = `${slotPath}.lock`;
+      const locked = await this.tryLock(lockPath);
+      if (!locked) {
+        continue;
       }
-      return output;
-    } catch (error) {
-      if (this.verbose) {
-        const message = error instanceof Error ? error.message : String(error);
-        console.log(
-          `[repo] git fail durationMs=${Date.now() - startedAt} args=${args.join(" ")} error=${message}`
-        );
+      const slotExists = (0, import_node_fs11.existsSync)(slotPath);
+      if (slotExists) {
+        await this.resetSlot(slotPath, templatePath, repos, poolReset);
+        return {
+          index: i,
+          path: slotPath,
+          isExisting: true,
+          lockPath,
+          fingerprint,
+          poolDir
+        };
       }
-      throw error;
+      await (0, import_promises27.mkdir)(slotPath, { recursive: true });
+      if (templatePath) {
+        await copyDirectoryRecursive2(templatePath, slotPath);
+      }
+      if (repos.length > 0) {
+        await repoManager.materializeAll(repos, slotPath);
+      }
+      await this.writeMetadata(poolDir, fingerprint, templatePath ?? null, repos);
+      return {
+        index: i,
+        path: slotPath,
+        isExisting: false,
+        lockPath,
+        fingerprint,
+        poolDir
+      };
+    }
+    throw new Error(
+      `All ${maxSlots} pool slots are locked for fingerprint ${fingerprint.slice(0, 12)}...`
+    );
+  }
+  /** Remove lock file to release a slot. */
+  async releaseSlot(slot) {
+    try {
+      await (0, import_promises27.unlink)(slot.lockPath);
+    } catch {
     }
   }
   /**
-   * Ensure a bare mirror cache exists for the given source.
-   * Creates on first access, fetches updates on subsequent calls.
-   * Returns the absolute path to the cache directory.
+   * Try to acquire a PID-based lock file.
+   * On EEXIST, read PID and check if process is alive. If dead, stale lock — remove and retry.
+   * Returns true if lock acquired, false if slot is actively locked.
+   * Uses a bounded loop (max 3 attempts) to avoid unbounded recursion.
    */
-  async ensureCache(source, depth, resolve) {
-    const key = cacheKey(source);
-    const cachePath = import_node_path39.default.join(this.cacheDir, key);
-    const lockPath = `${cachePath}.lock`;
-    const cacheExists = (0, import_node_fs11.existsSync)(import_node_path39.default.join(cachePath, "HEAD"));
-    if (this.verbose) {
-      console.log(
-        `[repo] ensureCache source=${getSourceUrl(source)} resolve=${resolve ?? "remote"} cache=${cachePath}`
-      );
+  async tryLock(lockPath) {
+    for (let attempt = 0; attempt < 3; attempt++) {
+      try {
+        await (0, import_promises27.writeFile)(lockPath, String(process.pid), { flag: "wx" });
+        return true;
+      } catch (err) {
+        if (err.code !== "EEXIST") {
+          throw err;
+        }
+        try {
+          const pidStr = await (0, import_promises27.readFile)(lockPath, "utf-8");
+          const pid = Number.parseInt(pidStr.trim(), 10);
+          if (!Number.isNaN(pid)) {
+            try {
+              process.kill(pid, 0);
+              return false;
+            } catch {
+              await (0, import_promises27.unlink)(lockPath).catch(() => {
+              });
+              continue;
+            }
+          }
+        } catch {
+        }
+        return false;
+      }
+    }
+    return false;
+  }
+  /**
+   * Check if the stored fingerprint in metadata.json differs from the computed one.
+   * Returns true if drifted, false otherwise.
+   * Returns false (no drift) if metadata.json doesn't exist (first use).
+   */
+  async checkDrift(poolDir, fingerprint) {
+    const metadataPath = import_node_path39.default.join(poolDir, "metadata.json");
+    try {
+      const raw = await (0, import_promises27.readFile)(metadataPath, "utf-8");
+      const metadata = JSON.parse(raw);
+      return metadata.fingerprint !== fingerprint;
+    } catch {
+      return false;
     }
-    if (resolve === "local") {
-      if (cacheExists) {
-        if (this.verbose) {
-          console.log(`[repo] using existing local cache ${cachePath}`);
+  }
+  /** Write metadata.json with fingerprint, inputs, and timestamp. */
+  async writeMetadata(poolDir, fingerprint, templatePath, repos) {
+    const metadata = {
+      fingerprint,
+      templatePath,
+      repos,
+      createdAt: (/* @__PURE__ */ new Date()).toISOString()
+    };
+    await (0, import_promises27.writeFile)(import_node_path39.default.join(poolDir, "metadata.json"), JSON.stringify(metadata, null, 2));
+  }
+  /** Remove all slot directories and their lock files from a pool directory. */
+  async removeAllSlots(poolDir) {
+    const entries = await (0, import_promises27.readdir)(poolDir);
+    for (const entry of entries) {
+      if (entry.startsWith("slot-") && !entry.endsWith(".lock")) {
+        const lockPath = import_node_path39.default.join(poolDir, `${entry}.lock`);
+        if ((0, import_node_fs11.existsSync)(lockPath)) {
+          try {
+            const pidStr = await (0, import_promises27.readFile)(lockPath, "utf-8");
+            const pid = Number.parseInt(pidStr.trim(), 10);
+            if (!Number.isNaN(pid)) {
+              try {
+                process.kill(pid, 0);
+                console.warn(`[workspace-pool] Skipping slot ${entry}: locked by PID ${pid}`);
+                continue;
+              } catch {
+              }
+            }
+          } catch {
+          }
         }
-        return cachePath;
+        await (0, import_promises27.rm)(import_node_path39.default.join(poolDir, entry), { recursive: true, force: true });
+        await (0, import_promises27.rm)(lockPath, { force: true }).catch(() => {
+        });
       }
-      const url = getSourceUrl(source);
-      throw new Error(
-        `No cache found for \`${url}\`. Run \`agentv cache add --url ${url} --from <local-path>\` to seed it.`
+    }
+    await (0, import_promises27.rm)(import_node_path39.default.join(poolDir, "metadata.json"), { force: true }).catch(() => {
+    });
+  }
+  /**
+   * Reset an existing slot for reuse:
+   * 1. Reset repos (git reset --hard {ref} && git clean -fd per repo)
+   * 2. Re-copy template files (skip repo directories)
+   */
+  async resetSlot(slotPath, templatePath, repos, poolReset = "fast") {
+    for (const repo of repos) {
+      const repoDir = import_node_path39.default.join(slotPath, repo.path);
+      if (!(0, import_node_fs11.existsSync)(repoDir)) {
+        continue;
+      }
+      if (poolReset === "none") {
+        continue;
+      }
+      const ref = repo.checkout?.ref ?? "HEAD";
+      await git(["reset", "--hard", ref], { cwd: repoDir });
+      const cleanFlag = poolReset === "strict" ? "-fdx" : "-fd";
+      await git(["clean", cleanFlag], { cwd: repoDir });
+    }
+    if (templatePath) {
+      const repoDirNames = new Set(
+        repos.map((r) => {
+          const normalized = r.path.replace(/^\.\//, "");
+          return normalized.split("/")[0];
+        })
       );
+      await copyDirectoryRecursive2(templatePath, slotPath, repoDirNames);
     }
-    await (0, import_promises27.mkdir)(this.cacheDir, { recursive: true });
-    const lockStartedAt = Date.now();
-    await acquireLock(lockPath);
+  }
+};
+
+// src/evaluation/workspace/repo-manager.ts
+var import_node_child_process8 = require("child_process");
+var import_node_path40 = __toESM(require("path"), 1);
+var import_node_util6 = require("util");
+var execFileAsync2 = (0, import_node_util6.promisify)(import_node_child_process8.execFile);
+var DEFAULT_TIMEOUT_MS2 = 3e5;
+function gitEnv2() {
+  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"
+  };
+}
+function getSourceUrl(source) {
+  return source.type === "git" ? source.url : source.path;
+}
+async function git2(args, opts) {
+  const { stdout } = await execFileAsync2("git", args, {
+    cwd: opts?.cwd,
+    timeout: opts?.timeout ?? DEFAULT_TIMEOUT_MS2,
+    env: gitEnv2(),
+    maxBuffer: 50 * 1024 * 1024
+    // 50MB
+  });
+  return stdout.trim();
+}
+var RepoManager = class {
+  verbose;
+  constructor(verbose = false) {
+    this.verbose = verbose;
+  }
+  async runGit(args, opts) {
+    const startedAt = Date.now();
     if (this.verbose) {
-      console.log(
-        `[repo] lock acquired path=${lockPath} waitedMs=${Date.now() - lockStartedAt}`
-      );
+      console.log(`[repo] git start cwd=${opts?.cwd ?? process.cwd()} args=${args.join(" ")}`);
     }
     try {
-      if (cacheExists) {
-        if (this.verbose) {
-          console.log(`[repo] refreshing existing cache ${cachePath}`);
-        }
-        const fetchArgs = ["fetch", "--prune"];
-        if (depth) {
-          fetchArgs.push("--depth", String(depth));
-        }
-        await this.runGit(fetchArgs, { cwd: cachePath });
-      } else {
-        if (this.verbose) {
-          console.log(`[repo] creating new cache ${cachePath}`);
-        }
-        const cloneArgs = ["clone", "--mirror", "--bare"];
-        if (depth) {
-          cloneArgs.push("--depth", String(depth));
-        }
-        const sourceUrl = getSourceUrl(source);
-        const cloneUrl = depth && source.type === "local" ? `file://${sourceUrl}` : sourceUrl;
-        cloneArgs.push(cloneUrl, cachePath);
-        await this.runGit(cloneArgs);
+      const output = await git2(args, opts);
+      if (this.verbose) {
+        console.log(`[repo] git ok durationMs=${Date.now() - startedAt} args=${args.join(" ")}`);
       }
-    } finally {
-      await releaseLock(lockPath);
+      return output;
+    } catch (error) {
       if (this.verbose) {
-        console.log(`[repo] lock released path=${lockPath}`);
+        const message = error instanceof Error ? error.message : String(error);
+        console.log(
+          `[repo] git fail durationMs=${Date.now() - startedAt} args=${args.join(" ")} error=${message}`
+        );
       }
+      throw error;
     }
-    return cachePath;
   }
   /**
-   * Clone a repo from cache into the workspace at the configured path.
+   * Clone a repo directly from source into the workspace at the configured path.
    * Handles checkout, ref resolution, ancestor walking, shallow clone, sparse checkout.
    */
   async materialize(repo, workspacePath) {
-    const targetDir = import_node_path39.default.join(workspacePath, repo.path);
+    const targetDir = import_node_path40.default.join(workspacePath, repo.path);
+    const sourceUrl = getSourceUrl(repo.source);
     const startedAt = Date.now();
     if (this.verbose) {
       console.log(
-        `[repo] materialize start path=${repo.path} source=${getSourceUrl(repo.source)} workspace=${workspacePath}`
+        `[repo] materialize start path=${repo.path} source=${sourceUrl} workspace=${workspacePath}`
       );
     }
-    const cachePath = await this.ensureCache(
-      repo.source,
-      repo.clone?.depth,
-      repo.checkout?.resolve
-    );
     const cloneArgs = ["clone"];
     if (repo.clone?.depth) {
       cloneArgs.push("--depth", String(repo.clone.depth));
@@ -15709,7 +16056,7 @@ var RepoManager = class {
       cloneArgs.push("--filter", repo.clone.filter);
     }
     cloneArgs.push("--no-checkout");
-    const cloneUrl = repo.clone?.depth || repo.clone?.filter ? `file://${cachePath}` : cachePath;
+    const cloneUrl = (repo.clone?.depth || repo.clone?.filter) && repo.source.type === "local" ? `file://${sourceUrl}` : sourceUrl;
     cloneArgs.push(cloneUrl, targetDir);
     await this.runGit(cloneArgs);
     if (repo.clone?.sparse?.length) {
@@ -15781,66 +16128,28 @@ var RepoManager = class {
     }
   }
   /** Reset repos in workspace to their checkout state. */
-  async reset(repos, workspacePath, strategy) {
-    if (strategy === "recreate") {
-      for (const repo of repos) {
-        const targetDir = import_node_path39.default.join(workspacePath, repo.path);
-        await (0, import_promises27.rm)(targetDir, { recursive: true, force: true });
-      }
-      await this.materializeAll(repos, workspacePath);
-      return;
-    }
+  async reset(repos, workspacePath, reset) {
+    const cleanFlag = reset === "strict" ? "-fdx" : "-fd";
     for (const repo of repos) {
-      const targetDir = import_node_path39.default.join(workspacePath, repo.path);
+      const targetDir = import_node_path40.default.join(workspacePath, repo.path);
       await this.runGit(["reset", "--hard", "HEAD"], { cwd: targetDir });
-      await this.runGit(["clean", "-fd"], { cwd: targetDir });
-    }
-  }
-  /**
-   * Seed the cache from a local repository, setting the remote to a given URL.
-   * Useful for avoiding slow network clones when a local clone already exists.
-   */
-  async seedCache(localPath, remoteUrl, opts) {
-    const source = { type: "git", url: remoteUrl };
-    const key = cacheKey(source);
-    const cachePath = import_node_path39.default.join(this.cacheDir, key);
-    const lockPath = `${cachePath}.lock`;
-    await (0, import_promises27.mkdir)(this.cacheDir, { recursive: true });
-    await acquireLock(lockPath);
-    try {
-      if ((0, import_node_fs11.existsSync)(import_node_path39.default.join(cachePath, "HEAD"))) {
-        if (!opts?.force) {
-          throw new Error(
-            `Cache already exists for ${remoteUrl} at ${cachePath}. Use force to overwrite.`
-          );
-        }
-        await (0, import_promises27.rm)(cachePath, { recursive: true, force: true });
-      }
-      await git(["clone", "--mirror", "--bare", localPath, cachePath]);
-      await git(["remote", "set-url", "origin", remoteUrl], { cwd: cachePath });
-    } finally {
-      await releaseLock(lockPath);
+      await this.runGit(["clean", cleanFlag], { cwd: targetDir });
     }
-    return cachePath;
-  }
-  /** Remove the entire cache directory. */
-  async cleanCache() {
-    await (0, import_promises27.rm)(this.cacheDir, { recursive: true, force: true });
   }
 };
 
 // src/evaluation/workspace/resolve.ts
 var import_promises28 = require("fs/promises");
-var import_node_path40 = __toESM(require("path"), 1);
+var import_node_path41 = __toESM(require("path"), 1);
 async function resolveWorkspaceTemplate(templatePath) {
   if (!templatePath) {
     return void 0;
   }
-  const resolved = import_node_path40.default.resolve(templatePath);
+  const resolved = import_node_path41.default.resolve(templatePath);
   const stats = await (0, import_promises28.stat)(resolved);
   if (stats.isFile()) {
     return {
-      dir: import_node_path40.default.dirname(resolved),
+      dir: import_node_path41.default.dirname(resolved),
       workspaceFile: resolved
     };
   }
@@ -15852,14 +16161,14 @@ async function resolveWorkspaceTemplate(templatePath) {
   if (workspaceFiles.length === 1) {
     return {
       dir: resolved,
-      workspaceFile: import_node_path40.default.join(resolved, workspaceFiles[0])
+      workspaceFile: import_node_path41.default.join(resolved, workspaceFiles[0])
     };
   }
   if (workspaceFiles.length > 1) {
     const conventionFile = workspaceFiles.find((f) => f === "template.code-workspace");
     return {
       dir: resolved,
-      workspaceFile: conventionFile ? import_node_path40.default.join(resolved, conventionFile) : void 0
+      workspaceFile: conventionFile ? import_node_path41.default.join(resolved, conventionFile) : void 0
     };
   }
   return { dir: resolved };
@@ -15911,6 +16220,22 @@ function classifyQualityStatus(score) {
 function usesFileReferencePrompt(provider) {
   return isAgentProvider(provider) || provider.kind === "cli";
 }
+function toScriptConfig(hook, hookName, context2) {
+  const command = hook.command ?? hook.script;
+  if (!command || command.length === 0) {
+    throw new Error(`${hookName} hook in ${context2} 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 }
+  };
+}
+function hasHookCommand(hook) {
+  return !!(hook?.command && hook.command.length > 0 || hook?.script && hook.script.length > 0);
+}
 function getWorkspaceTemplate(target) {
   const config = target.config;
   if ("workspaceTemplate" in config && typeof config.workspaceTemplate === "string") {
@@ -15941,7 +16266,15 @@ async function runEvaluation(options) {
     trials,
     streamCallbacks,
     totalBudgetUsd,
-    failOnError
+    failOnError,
+    poolWorkspaces,
+    poolMaxSlots: configPoolMaxSlots,
+    workspace: legacyWorkspacePath,
+    workspaceMode,
+    workspacePath,
+    workspaceClean,
+    retainOnSuccess,
+    retainOnFailure
   } = options;
   let useCache = options.useCache;
   if (trials && trials.count > 1 && useCache) {
@@ -16015,7 +16348,7 @@ async function runEvaluation(options) {
   ];
   const evaluatorRegistry = buildEvaluatorRegistry(evaluators, resolveJudgeProvider);
   const typeRegistry = createBuiltinRegistry();
-  const discoveryBaseDir = evalFilePath ? import_node_path41.default.dirname(import_node_path41.default.resolve(evalFilePath)) : process.cwd();
+  const discoveryBaseDir = evalFilePath ? import_node_path42.default.dirname(import_node_path42.default.resolve(evalFilePath)) : process.cwd();
   const evalDir = discoveryBaseDir;
   await discoverAssertions(typeRegistry, discoveryBaseDir);
   const providerRegistry = createBuiltinProviderRegistry();
@@ -16077,13 +16410,29 @@ async function runEvaluation(options) {
     }
   };
   const isPerTestIsolation = suiteWorkspace?.isolation === "per_test";
-  const hasSharedWorkspace = !!(workspaceTemplate || suiteWorkspace?.before_all || suiteWorkspace?.repos?.length && !isPerTestIsolation);
+  const configuredMode = suiteWorkspace?.mode ?? workspaceMode;
+  const configuredStaticPath = suiteWorkspace?.static_path ?? workspacePath ?? legacyWorkspacePath;
+  const useStaticWorkspace = configuredMode === "static" || !!configuredStaticPath && !configuredMode;
+  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.mode=static requires workspace.static_path or --workspace-path");
+  }
+  const hasSharedWorkspace = !!(useStaticWorkspace || workspaceTemplate || suiteWorkspace?.hooks || suiteWorkspace?.repos?.length && !isPerTestIsolation);
+  const poolEnabled = configuredMode === "pooled" ? true : configuredMode === "ephemeral" || useStaticWorkspace ? false : suiteWorkspace?.pool ?? poolWorkspaces ?? true;
+  const usePool = poolEnabled !== false && !!suiteWorkspace?.repos?.length && !isPerTestIsolation && !useStaticWorkspace;
+  const finishCleanPolicy = suiteWorkspace?.hooks?.on_finish?.clean;
+  const resolvedRetainOnSuccess = (finishCleanPolicy === "always" || finishCleanPolicy === "on_success" ? "cleanup" : finishCleanPolicy === "on_failure" || finishCleanPolicy === "never" ? "keep" : void 0) ?? retainOnSuccess ?? (keepWorkspaces ? "keep" : "cleanup");
+  const resolvedRetainOnFailure = (finishCleanPolicy === "always" || finishCleanPolicy === "on_failure" ? "cleanup" : finishCleanPolicy === "on_success" || finishCleanPolicy === "never" ? "keep" : void 0) ?? retainOnFailure ?? (cleanupWorkspaces ? "cleanup" : "keep");
   const requestedWorkers = options.maxConcurrency ?? target.workers ?? 1;
-  const workers = hasSharedWorkspace ? 1 : requestedWorkers;
+  const workers = hasSharedWorkspace && !usePool ? 1 : requestedWorkers;
   setupLog(
-    `sharedWorkspace=${hasSharedWorkspace} perTestIsolation=${isPerTestIsolation} requestedWorkers=${requestedWorkers} effectiveWorkers=${workers}`
+    `sharedWorkspace=${hasSharedWorkspace} perTestIsolation=${isPerTestIsolation} usePool=${usePool} requestedWorkers=${requestedWorkers} effectiveWorkers=${workers}`
   );
-  if (hasSharedWorkspace && requestedWorkers > 1) {
+  if (hasSharedWorkspace && !usePool && requestedWorkers > 1) {
     console.warn(
       `Warning: Shared workspace requires sequential execution. Overriding workers from ${requestedWorkers} to 1.`
     );
@@ -16092,7 +16441,38 @@ async function runEvaluation(options) {
   let sharedWorkspacePath;
   let sharedBaselineCommit;
   let beforeAllOutput;
-  if (workspaceTemplate) {
+  let poolManager;
+  let poolSlot;
+  const poolSlots = [];
+  const availablePoolSlots = [];
+  const poolSlotBaselines = /* @__PURE__ */ new Map();
+  const poolMaxSlots = Math.min(configPoolMaxSlots ?? 10, 50);
+  if (useStaticWorkspace && configuredStaticPath) {
+    sharedWorkspacePath = configuredStaticPath;
+    setupLog(`using static workspace: ${configuredStaticPath}`);
+  } else if (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) ?? suiteWorkspace.hooks?.on_reuse?.reset ?? "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 (workspaceTemplate) {
     setupLog(`creating shared workspace from template: ${workspaceTemplate}`);
     try {
       sharedWorkspacePath = await createTempWorkspace(workspaceTemplate, evalRunId, "shared");
@@ -16101,288 +16481,359 @@ async function runEvaluation(options) {
       const message = error instanceof Error ? error.message : String(error);
       throw new Error(`Failed to create shared workspace: ${message}`);
     }
+  } else if (suiteWorkspace?.hooks || suiteWorkspace?.repos?.length && !isPerTestIsolation) {
+    sharedWorkspacePath = getWorkspacePath(evalRunId, "shared");
+    await (0, import_promises29.mkdir)(sharedWorkspacePath, { recursive: true });
+    setupLog(`created empty shared workspace at: ${sharedWorkspacePath}`);
+  }
+  try {
     if (suiteWorkspaceFile && sharedWorkspacePath) {
-      const copiedWorkspaceFile = import_node_path41.default.join(sharedWorkspacePath, import_node_path41.default.basename(suiteWorkspaceFile));
+      const copiedWorkspaceFile = import_node_path42.default.join(sharedWorkspacePath, import_node_path42.default.basename(suiteWorkspaceFile));
       try {
         await (0, import_promises29.stat)(copiedWorkspaceFile);
         suiteWorkspaceFile = copiedWorkspaceFile;
       } catch {
       }
     }
-  } else if (suiteWorkspace?.before_all || suiteWorkspace?.repos?.length && !isPerTestIsolation) {
-    sharedWorkspacePath = getWorkspacePath(evalRunId, "shared");
-    await (0, import_promises29.mkdir)(sharedWorkspacePath, { recursive: true });
-    setupLog(`created empty shared workspace at: ${sharedWorkspacePath}`);
-  }
-  const repoManager = suiteWorkspace?.repos?.length ? new RepoManager(void 0, verbose) : void 0;
-  if (repoManager && sharedWorkspacePath && suiteWorkspace?.repos && !isPerTestIsolation) {
-    setupLog(`materializing ${suiteWorkspace.repos.length} shared repo(s) into ${sharedWorkspacePath}`);
-    try {
-      await repoManager.materializeAll(suiteWorkspace.repos, sharedWorkspacePath);
-      setupLog("shared repo materialization complete");
-    } catch (error) {
-      const message = error instanceof Error ? error.message : String(error);
-      if (sharedWorkspacePath) {
-        await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-        });
-      }
-      throw new Error(`Failed to materialize repos: ${message}`);
-    }
-  }
-  if (sharedWorkspacePath && suiteWorkspace?.before_all) {
-    const beforeAllCommand = (suiteWorkspace.before_all.command ?? suiteWorkspace.before_all.script ?? []).join(" ");
-    setupLog(
-      `running shared before_all in cwd=${suiteWorkspace.before_all.cwd ?? evalDir} command=${beforeAllCommand}`
-    );
-    const scriptContext = {
-      workspacePath: sharedWorkspacePath,
-      testId: "__before_all__",
-      evalRunId,
-      evalDir
-    };
-    try {
-      beforeAllOutput = await executeWorkspaceScript(suiteWorkspace.before_all, scriptContext);
-      setupLog("shared before_all completed");
-    } catch (error) {
-      const message = error instanceof Error ? error.message : String(error);
-      if (sharedWorkspacePath) {
-        await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-        });
-      }
-      throw new Error(`before_all script failed: ${message}`);
-    }
-  }
-  if (sharedWorkspacePath) {
-    try {
-      sharedBaselineCommit = await initializeBaseline(sharedWorkspacePath);
-      setupLog(`shared baseline initialized: ${sharedBaselineCommit}`);
-    } catch {
-      setupLog("shared baseline initialization skipped (non-fatal)");
-    }
-  }
-  let nextWorkerId = 1;
-  const workerIdByEvalId = /* @__PURE__ */ new Map();
-  let beforeAllOutputAttached = false;
-  let cumulativeBudgetCost = 0;
-  let budgetExhausted = false;
-  let failOnErrorTriggered = false;
-  const promises = filteredEvalCases.map(
-    (evalCase) => limit(async () => {
-      const workerId = nextWorkerId++;
-      workerIdByEvalId.set(evalCase.id, workerId);
-      if (totalBudgetUsd !== void 0 && budgetExhausted) {
-        const budgetResult = {
-          timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
-          testId: evalCase.id,
-          dataset: evalCase.dataset,
-          score: 0,
-          hits: [],
-          misses: [],
-          answer: "",
-          target: target.name,
-          error: `Suite budget exceeded ($${cumulativeBudgetCost.toFixed(4)} / $${totalBudgetUsd.toFixed(4)})`,
-          budgetExceeded: true,
-          executionStatus: "execution_error",
-          failureStage: "setup",
-          failureReasonCode: "budget_exceeded",
-          executionError: {
-            message: `Suite budget exceeded ($${cumulativeBudgetCost.toFixed(4)} / $${totalBudgetUsd.toFixed(4)})`,
-            stage: "setup"
-          }
-        };
-        if (onProgress) {
-          await onProgress({
-            workerId,
-            testId: evalCase.id,
-            status: "failed",
-            completedAt: Date.now(),
-            error: budgetResult.error
+    const repoManager = suiteWorkspace?.repos?.length && !usePool && !useStaticWorkspace ? new RepoManager(verbose) : void 0;
+    if (repoManager && sharedWorkspacePath && suiteWorkspace?.repos && !isPerTestIsolation) {
+      setupLog(
+        `materializing ${suiteWorkspace.repos.length} shared repo(s) into ${sharedWorkspacePath}`
+      );
+      try {
+        await repoManager.materializeAll(suiteWorkspace.repos, sharedWorkspacePath);
+        setupLog("shared repo materialization complete");
+      } catch (error) {
+        const message = error instanceof Error ? error.message : String(error);
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
           });
         }
-        if (onResult) {
-          await onResult(budgetResult);
-        }
-        return budgetResult;
+        throw new Error(`Failed to materialize repos: ${message}`);
       }
-      if (failOnError === true && failOnErrorTriggered) {
-        const errorMsg = "Halted: execution error encountered with fail_on_error enabled";
-        const haltResult = {
-          timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
-          testId: evalCase.id,
-          dataset: evalCase.dataset,
-          score: 0,
-          hits: [],
-          misses: [],
-          answer: "",
-          target: target.name,
-          error: errorMsg,
-          executionStatus: "execution_error",
-          failureStage: "setup",
-          failureReasonCode: "error_threshold_exceeded",
-          executionError: { message: errorMsg, stage: "setup" }
-        };
-        if (onProgress) {
-          await onProgress({
-            workerId,
-            testId: evalCase.id,
-            status: "failed",
-            completedAt: Date.now(),
-            error: haltResult.error
+    }
+    const suiteBeforeAllHook = suiteWorkspace?.hooks?.before_all_tests;
+    if (sharedWorkspacePath && hasHookCommand(suiteBeforeAllHook)) {
+      const beforeAllHook = suiteBeforeAllHook;
+      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
+      setupLog(
+        `running shared before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
+      );
+      const scriptContext = {
+        workspacePath: sharedWorkspacePath,
+        testId: "__before_all__",
+        evalRunId,
+        evalDir
+      };
+      try {
+        beforeAllOutput = await executeWorkspaceScript(
+          toScriptConfig(beforeAllHook, "before_all_tests", "suite workspace"),
+          scriptContext
+        );
+        setupLog("shared before_all completed");
+      } catch (error) {
+        const message = error instanceof Error ? error.message : String(error);
+        if (sharedWorkspacePath && !useStaticWorkspace) {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
           });
         }
-        if (onResult) {
-          await onResult(haltResult);
-        }
-        return haltResult;
-      }
-      if (onProgress) {
-        await onProgress({
-          workerId,
-          testId: evalCase.id,
-          status: "running",
-          startedAt: Date.now()
-        });
+        throw new Error(`before_all script failed: ${message}`);
       }
-      try {
-        const judgeProvider = await resolveJudgeProvider(target);
-        const runCaseOptions = {
-          evalCase,
-          provider: primaryProvider,
-          target,
-          evaluators: evaluatorRegistry,
-          maxRetries,
-          agentTimeoutMs,
-          cache,
-          useCache,
-          now,
-          judgeProvider,
-          targetResolver,
-          availableTargets,
+    }
+    if (availablePoolSlots.length > 0 && 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,
-          keepWorkspaces,
-          cleanupWorkspaces,
-          sharedWorkspacePath,
-          sharedBaselineCommit,
-          suiteWorkspaceFile,
-          streamCallbacks,
-          typeRegistry,
-          repoManager,
           evalDir
         };
-        let result = trials && trials.count > 1 ? await runEvalCaseWithTrials(runCaseOptions, trials) : await runEvalCase(runCaseOptions);
-        if (totalBudgetUsd !== void 0) {
-          let caseCost;
-          if (result.trials && result.trials.length > 0) {
-            const trialCostSum = result.trials.reduce((sum, t) => sum + (t.costUsd ?? 0), 0);
-            if (trialCostSum > 0) {
-              caseCost = trialCostSum;
+        try {
+          const output = await executeWorkspaceScript(
+            toScriptConfig(beforeAllHook, "before_all_tests", "suite workspace"),
+            scriptContext
+          );
+          if (!beforeAllOutput) beforeAllOutput = output;
+          setupLog(`before_all completed on pool slot ${slot.index}`);
+        } catch (error) {
+          const message = error instanceof Error ? error.message : String(error);
+          throw new Error(`before_all script failed on pool slot ${slot.index}: ${message}`);
+        }
+      }
+    }
+    if (sharedWorkspacePath) {
+      try {
+        sharedBaselineCommit = await initializeBaseline(sharedWorkspacePath);
+        setupLog(`shared baseline initialized: ${sharedBaselineCommit}`);
+      } catch {
+        setupLog("shared baseline initialization skipped (non-fatal)");
+      }
+    }
+    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 {
+          setupLog(`pool slot ${slot.index} baseline initialization skipped (non-fatal)`);
+        }
+      }
+    }
+    let nextWorkerId = 1;
+    const workerIdByEvalId = /* @__PURE__ */ new Map();
+    let beforeAllOutputAttached = false;
+    let cumulativeBudgetCost = 0;
+    let budgetExhausted = false;
+    let failOnErrorTriggered = false;
+    const promises = filteredEvalCases.map(
+      (evalCase) => limit(async () => {
+        const workerId = nextWorkerId++;
+        workerIdByEvalId.set(evalCase.id, workerId);
+        if (totalBudgetUsd !== void 0 && budgetExhausted) {
+          const budgetResult = {
+            timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
+            testId: evalCase.id,
+            dataset: evalCase.dataset,
+            score: 0,
+            hits: [],
+            misses: [],
+            answer: "",
+            target: target.name,
+            error: `Suite budget exceeded ($${cumulativeBudgetCost.toFixed(4)} / $${totalBudgetUsd.toFixed(4)})`,
+            budgetExceeded: true,
+            executionStatus: "execution_error",
+            failureStage: "setup",
+            failureReasonCode: "budget_exceeded",
+            executionError: {
+              message: `Suite budget exceeded ($${cumulativeBudgetCost.toFixed(4)} / $${totalBudgetUsd.toFixed(4)})`,
+              stage: "setup"
             }
-          } else {
-            caseCost = result.costUsd;
+          };
+          if (onProgress) {
+            await onProgress({
+              workerId,
+              testId: evalCase.id,
+              status: "failed",
+              completedAt: Date.now(),
+              error: budgetResult.error
+            });
           }
-          if (caseCost !== void 0) {
-            cumulativeBudgetCost += caseCost;
-            if (cumulativeBudgetCost >= totalBudgetUsd) {
-              budgetExhausted = true;
-            }
+          if (onResult) {
+            await onResult(budgetResult);
           }
+          return budgetResult;
         }
-        if (failOnError === true && result.executionStatus === "execution_error") {
-          failOnErrorTriggered = true;
-        }
-        if (beforeAllOutput && !beforeAllOutputAttached) {
-          result = { ...result, beforeAllOutput };
-          beforeAllOutputAttached = true;
+        if (failOnError === true && failOnErrorTriggered) {
+          const errorMsg = "Halted: execution error encountered with fail_on_error enabled";
+          const haltResult = {
+            timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
+            testId: evalCase.id,
+            dataset: evalCase.dataset,
+            score: 0,
+            hits: [],
+            misses: [],
+            answer: "",
+            target: target.name,
+            error: errorMsg,
+            executionStatus: "execution_error",
+            failureStage: "setup",
+            failureReasonCode: "error_threshold_exceeded",
+            executionError: { message: errorMsg, stage: "setup" }
+          };
+          if (onProgress) {
+            await onProgress({
+              workerId,
+              testId: evalCase.id,
+              status: "failed",
+              completedAt: Date.now(),
+              error: haltResult.error
+            });
+          }
+          if (onResult) {
+            await onResult(haltResult);
+          }
+          return haltResult;
         }
         if (onProgress) {
           await onProgress({
             workerId,
             testId: evalCase.id,
-            status: result.error ? "failed" : "completed",
-            startedAt: 0,
-            // Not used for completed status
-            completedAt: Date.now(),
-            error: result.error
+            status: "running",
+            startedAt: Date.now()
           });
         }
-        if (onResult) {
-          await onResult(result);
+        const testPoolSlot = availablePoolSlots.length > 0 ? availablePoolSlots.pop() : void 0;
+        const testWorkspacePath = testPoolSlot?.path ?? sharedWorkspacePath;
+        const testBaselineCommit = testPoolSlot ? poolSlotBaselines.get(testPoolSlot.path) : sharedBaselineCommit;
+        try {
+          const judgeProvider = await resolveJudgeProvider(target);
+          const runCaseOptions = {
+            evalCase,
+            provider: primaryProvider,
+            target,
+            evaluators: evaluatorRegistry,
+            maxRetries,
+            agentTimeoutMs,
+            cache,
+            useCache,
+            now,
+            judgeProvider,
+            targetResolver,
+            availableTargets,
+            evalRunId,
+            keepWorkspaces,
+            cleanupWorkspaces,
+            retainOnSuccess: resolvedRetainOnSuccess,
+            retainOnFailure: resolvedRetainOnFailure,
+            sharedWorkspacePath: testWorkspacePath,
+            sharedBaselineCommit: testBaselineCommit,
+            suiteWorkspaceFile,
+            streamCallbacks,
+            typeRegistry,
+            repoManager,
+            evalDir
+          };
+          let result = trials && trials.count > 1 ? await runEvalCaseWithTrials(runCaseOptions, trials) : await runEvalCase(runCaseOptions);
+          if (totalBudgetUsd !== void 0) {
+            let caseCost;
+            if (result.trials && result.trials.length > 0) {
+              const trialCostSum = result.trials.reduce((sum, t) => sum + (t.costUsd ?? 0), 0);
+              if (trialCostSum > 0) {
+                caseCost = trialCostSum;
+              }
+            } else {
+              caseCost = result.costUsd;
+            }
+            if (caseCost !== void 0) {
+              cumulativeBudgetCost += caseCost;
+              if (cumulativeBudgetCost >= totalBudgetUsd) {
+                budgetExhausted = true;
+              }
+            }
+          }
+          if (failOnError === true && result.executionStatus === "execution_error") {
+            failOnErrorTriggered = true;
+          }
+          if (beforeAllOutput && !beforeAllOutputAttached) {
+            result = { ...result, beforeAllOutput };
+            beforeAllOutputAttached = true;
+          }
+          if (onProgress) {
+            await onProgress({
+              workerId,
+              testId: evalCase.id,
+              status: result.error ? "failed" : "completed",
+              startedAt: 0,
+              // Not used for completed status
+              completedAt: Date.now(),
+              error: result.error
+            });
+          }
+          if (onResult) {
+            await onResult(result);
+          }
+          return result;
+        } catch (error) {
+          if (onProgress) {
+            await onProgress({
+              workerId,
+              testId: evalCase.id,
+              status: "failed",
+              completedAt: Date.now(),
+              error: error instanceof Error ? error.message : String(error)
+            });
+          }
+          throw error;
+        } finally {
+          if (testPoolSlot) {
+            availablePoolSlots.push(testPoolSlot);
+          }
         }
-        return result;
-      } catch (error) {
-        if (onProgress) {
-          await onProgress({
-            workerId,
-            testId: evalCase.id,
-            status: "failed",
-            completedAt: Date.now(),
-            error: error instanceof Error ? error.message : String(error)
-          });
+      })
+    );
+    const settled = await Promise.allSettled(promises);
+    const results = [];
+    for (let i = 0; i < settled.length; i++) {
+      const outcome = settled[i];
+      if (outcome.status === "fulfilled") {
+        results.push(outcome.value);
+      } else {
+        const evalCase = filteredEvalCases[i];
+        const formattingMode = usesFileReferencePrompt(primaryProvider) ? "agent" : "lm";
+        const promptInputs = await buildPromptInputs(evalCase, formattingMode);
+        const errorResult = buildErrorResult(
+          evalCase,
+          target.name,
+          (now ?? (() => /* @__PURE__ */ new Date()))(),
+          outcome.reason,
+          promptInputs,
+          primaryProvider,
+          "agent",
+          "provider_error"
+        );
+        results.push(errorResult);
+        if (onResult) {
+          await onResult(errorResult);
         }
-        throw error;
       }
-    })
-  );
-  const settled = await Promise.allSettled(promises);
-  const results = [];
-  for (let i = 0; i < settled.length; i++) {
-    const outcome = settled[i];
-    if (outcome.status === "fulfilled") {
-      results.push(outcome.value);
-    } else {
-      const evalCase = filteredEvalCases[i];
-      const formattingMode = usesFileReferencePrompt(primaryProvider) ? "agent" : "lm";
-      const promptInputs = await buildPromptInputs(evalCase, formattingMode);
-      const errorResult = buildErrorResult(
-        evalCase,
-        target.name,
-        (now ?? (() => /* @__PURE__ */ new Date()))(),
-        outcome.reason,
-        promptInputs,
-        primaryProvider,
-        "agent",
-        "provider_error"
-      );
-      results.push(errorResult);
-      if (onResult) {
-        await onResult(errorResult);
+    }
+    const afterAllWorkspaces = poolSlots.length > 1 ? poolSlots.map((s) => s.path) : sharedWorkspacePath ? [sharedWorkspacePath] : [];
+    const suiteAfterAllHook = suiteWorkspace?.hooks?.after_all_tests;
+    if (afterAllWorkspaces.length > 0 && hasHookCommand(suiteAfterAllHook)) {
+      const afterAllHook = suiteAfterAllHook;
+      for (const wsPath of afterAllWorkspaces) {
+        const scriptContext = {
+          workspacePath: wsPath,
+          testId: "__after_all__",
+          evalRunId,
+          evalDir
+        };
+        try {
+          const afterAllOutput = await executeWorkspaceScript(
+            toScriptConfig(afterAllHook, "after_all_tests", "suite workspace"),
+            scriptContext,
+            "warn"
+          );
+          if (afterAllOutput && results.length > 0 && wsPath === afterAllWorkspaces[0]) {
+            results[results.length - 1] = { ...results[results.length - 1], afterAllOutput };
+          }
+        } catch {
+        }
       }
     }
-  }
-  if (sharedWorkspacePath && suiteWorkspace?.after_all) {
-    const scriptContext = {
-      workspacePath: sharedWorkspacePath,
-      testId: "__after_all__",
-      evalRunId,
-      evalDir
-    };
-    try {
-      const afterAllOutput = await executeWorkspaceScript(
-        suiteWorkspace.after_all,
-        scriptContext,
-        "warn"
-      );
-      if (afterAllOutput && results.length > 0) {
-        results[results.length - 1] = { ...results[results.length - 1], afterAllOutput };
+    if (sharedWorkspacePath && !poolSlot && poolSlots.length === 0 && !useStaticWorkspace) {
+      const hasFailure = results.some((r) => !!r.error || r.score < 0.5);
+      if (hasFailure) {
+        if (resolvedRetainOnFailure === "cleanup") {
+          await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+          });
+        }
+      } else if (resolvedRetainOnSuccess === "cleanup") {
+        await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+        });
       }
-    } catch {
     }
-  }
-  if (sharedWorkspacePath) {
-    const hasFailure = results.some((r) => !!r.error || r.score < 0.5);
     if (cleanupWorkspaces) {
-      await cleanupWorkspace(sharedWorkspacePath).catch(() => {
-      });
-    } else if (!hasFailure && !keepWorkspaces) {
-      await cleanupWorkspace(sharedWorkspacePath).catch(() => {
+      await cleanupEvalWorkspaces(evalRunId).catch(() => {
       });
     }
+    return results;
+  } finally {
+    if (poolManager) {
+      if (poolSlot) {
+        await poolManager.releaseSlot(poolSlot);
+      }
+      for (const slot of poolSlots) {
+        if (slot !== poolSlot) {
+          await poolManager.releaseSlot(slot).catch(() => {
+          });
+        }
+      }
+    }
   }
-  if (cleanupWorkspaces) {
-    await cleanupEvalWorkspaces(evalRunId).catch(() => {
-    });
-  }
-  return results;
 }
 async function runBatchEvaluation(options) {
   const {
@@ -16554,6 +17005,8 @@ async function runEvalCase(options) {
     evalRunId,
     keepWorkspaces,
     cleanupWorkspaces: forceCleanup,
+    retainOnSuccess,
+    retainOnFailure,
     sharedWorkspacePath,
     sharedBaselineCommit,
     suiteWorkspaceFile,
@@ -16565,10 +17018,10 @@ async function runEvalCase(options) {
   const formattingMode = usesFileReferencePrompt(provider) ? "agent" : "lm";
   const promptInputs = await buildPromptInputs(evalCase, formattingMode);
   const typeRegistry = providedTypeRegistry ?? createBuiltinRegistry();
-  const cacheKey2 = useCache ? createCacheKey(provider, target, evalCase, promptInputs) : void 0;
+  const cacheKey = useCache ? createCacheKey(provider, target, evalCase, promptInputs) : void 0;
   let cachedResponse;
-  if (cacheKey2 && cache) {
-    cachedResponse = await cache.get(cacheKey2);
+  if (cacheKey && cache) {
+    cachedResponse = await cache.get(cacheKey);
   }
   const nowFn = now ?? (() => /* @__PURE__ */ new Date());
   let workspacePath = sharedWorkspacePath;
@@ -16599,7 +17052,7 @@ async function runEvalCase(options) {
         );
       }
       if (caseWorkspaceFile && workspacePath) {
-        const copiedFile = import_node_path41.default.join(workspacePath, import_node_path41.default.basename(caseWorkspaceFile));
+        const copiedFile = import_node_path42.default.join(workspacePath, import_node_path42.default.basename(caseWorkspaceFile));
         try {
           await (0, import_promises29.stat)(copiedFile);
           caseWorkspaceFile = copiedFile;
@@ -16607,12 +17060,12 @@ async function runEvalCase(options) {
         }
       }
     }
-    if (!workspacePath && (evalCase.workspace?.before_all || evalCase.workspace?.repos?.length) && evalRunId) {
+    if (!workspacePath && (evalCase.workspace?.hooks || evalCase.workspace?.repos?.length) && evalRunId) {
       workspacePath = getWorkspacePath(evalRunId, evalCase.id);
       await (0, import_promises29.mkdir)(workspacePath, { recursive: true });
     }
     if (evalCase.workspace?.repos?.length && workspacePath) {
-      const perCaseRepoManager = new RepoManager(void 0, setupDebug);
+      const perCaseRepoManager = new RepoManager(setupDebug);
       try {
         if (setupDebug) {
           console.log(
@@ -16637,11 +17090,13 @@ async function runEvalCase(options) {
         );
       }
     }
-    if (workspacePath && evalCase.workspace?.before_all) {
-      const beforeAllCommand = (evalCase.workspace.before_all.command ?? evalCase.workspace.before_all.script ?? []).join(" ");
+    const caseBeforeAllHook = evalCase.workspace?.hooks?.before_all_tests;
+    if (workspacePath && hasHookCommand(caseBeforeAllHook)) {
+      const beforeAllHook = caseBeforeAllHook;
+      const beforeAllCommand = (beforeAllHook.command ?? beforeAllHook.script ?? []).join(" ");
       if (setupDebug) {
         console.log(
-          `[setup] test=${evalCase.id} running before_all in cwd=${evalCase.workspace.before_all.cwd ?? evalDir} command=${beforeAllCommand}`
+          `[setup] test=${evalCase.id} running before_all in cwd=${beforeAllHook.cwd ?? evalDir} command=${beforeAllCommand}`
         );
       }
       const scriptContext = {
@@ -16654,7 +17109,7 @@ async function runEvalCase(options) {
       };
       try {
         beforeAllOutput = await executeWorkspaceScript(
-          evalCase.workspace.before_all,
+          toScriptConfig(beforeAllHook, "before_all_tests", `test '${evalCase.id}'`),
           scriptContext
         );
         if (setupDebug) {
@@ -16679,7 +17134,9 @@ async function runEvalCase(options) {
       }
     }
   }
-  if (workspacePath && evalCase.workspace?.before_each) {
+  const caseBeforeEachHook = evalCase.workspace?.hooks?.before_each_test;
+  if (workspacePath && hasHookCommand(caseBeforeEachHook)) {
+    const beforeEachHook = caseBeforeEachHook;
     const scriptContext = {
       workspacePath,
       testId: evalCase.id,
@@ -16690,7 +17147,7 @@ async function runEvalCase(options) {
     };
     try {
       beforeEachOutput = await executeWorkspaceScript(
-        evalCase.workspace.before_each,
+        toScriptConfig(beforeEachHook, "before_each_test", `test '${evalCase.id}'`),
         scriptContext
       );
     } catch (error) {
@@ -16778,8 +17235,8 @@ async function runEvalCase(options) {
     }
     return errorResult;
   }
-  if (cacheKey2 && cache && !cachedResponse) {
-    await cache.set(cacheKey2, providerResponse);
+  if (cacheKey && cache && !cachedResponse) {
+    await cache.set(cacheKey, providerResponse);
   }
   const output = providerResponse.output;
   const hasExecutionMetrics = providerResponse.tokenUsage !== void 0 || providerResponse.costUsd !== void 0 || providerResponse.durationMs !== void 0;
@@ -16807,17 +17264,19 @@ async function runEvalCase(options) {
     }
   }
   const providerError = extractProviderError(providerResponse);
-  if (repoManager && workspacePath && evalCase.workspace?.reset?.after_each && evalCase.workspace.reset.strategy && evalCase.workspace.reset.strategy !== "none" && evalCase.workspace.repos) {
+  if (repoManager && workspacePath && evalCase.workspace?.hooks?.after_each_test?.reset && evalCase.workspace.hooks.after_each_test.reset !== "none" && evalCase.workspace.repos) {
     try {
       await repoManager.reset(
         evalCase.workspace.repos,
         workspacePath,
-        evalCase.workspace.reset.strategy
+        evalCase.workspace.hooks.after_each_test.reset
       );
     } catch {
     }
   }
-  if (workspacePath && evalCase.workspace?.after_each) {
+  const caseAfterEachHook = evalCase.workspace?.hooks?.after_each_test;
+  if (workspacePath && hasHookCommand(caseAfterEachHook)) {
+    const afterEachHook = caseAfterEachHook;
     const scriptContext = {
       workspacePath,
       testId: evalCase.id,
@@ -16828,7 +17287,7 @@ async function runEvalCase(options) {
     };
     try {
       afterEachOutput = await executeWorkspaceScript(
-        evalCase.workspace.after_each,
+        toScriptConfig(afterEachHook, "after_each_test", `test '${evalCase.id}'`),
         scriptContext,
         "warn"
       );
@@ -16878,8 +17337,13 @@ async function runEvalCase(options) {
         await cleanupWorkspace(workspacePath).catch(() => {
         });
       } else if (isFailure) {
-        return { ...finalResult, workspacePath };
-      } else if (!keepWorkspaces) {
+        if ((retainOnFailure ?? "keep") === "cleanup") {
+          await cleanupWorkspace(workspacePath).catch(() => {
+          });
+        } else {
+          return { ...finalResult, workspacePath };
+        }
+      } else if ((retainOnSuccess ?? (keepWorkspaces ? "keep" : "cleanup")) !== "keep") {
         await cleanupWorkspace(workspacePath).catch(() => {
         });
       }
@@ -16897,11 +17361,12 @@ async function runEvalCase(options) {
       "evaluator_error"
     );
     if (workspacePath && !isSharedWorkspace) {
-      if (forceCleanup) {
+      if (forceCleanup || (retainOnFailure ?? "keep") === "cleanup") {
         await cleanupWorkspace(workspacePath).catch(() => {
         });
+      } else {
+        return { ...errorResult, workspacePath, beforeEachOutput, afterEachOutput };
       }
-      return { ...errorResult, workspacePath, beforeEachOutput, afterEachOutput };
     }
     return { ...errorResult, beforeEachOutput, afterEachOutput };
   }
@@ -16920,7 +17385,9 @@ async function runEvalCaseWithTrials(options, trialsConfig) {
       useCache: false,
       // Force cleanup for intermediate trials
       cleanupWorkspaces: isLastDeclaredTrial ? options.cleanupWorkspaces : true,
-      keepWorkspaces: isLastDeclaredTrial ? options.keepWorkspaces : false
+      keepWorkspaces: isLastDeclaredTrial ? options.keepWorkspaces : false,
+      retainOnSuccess: isLastDeclaredTrial ? options.retainOnSuccess : "cleanup",
+      retainOnFailure: isLastDeclaredTrial ? options.retainOnFailure : "cleanup"
     };
     const result = await runEvalCase(trialOptions);
     allResults.push(result);
@@ -17209,7 +17676,7 @@ async function runEvaluatorList(options) {
     fileChanges,
     workspacePath
   };
-  const evalFileDir = evalCase.guideline_paths[0] ? import_node_path41.default.dirname(evalCase.guideline_paths[0]) : process.cwd();
+  const evalFileDir = evalCase.guideline_paths[0] ? import_node_path42.default.dirname(evalCase.guideline_paths[0]) : process.cwd();
   const dispatchContext = {
     judgeProvider,
     targetResolver,
@@ -17512,7 +17979,7 @@ function computeWeightedMean(entries) {
 
 // src/evaluation/evaluate.ts
 var import_node_fs12 = require("fs");
-var import_node_path42 = __toESM(require("path"), 1);
+var import_node_path43 = __toESM(require("path"), 1);
 async function evaluate(config) {
   const startTime = Date.now();
   if (config.tests && config.specFile) {
@@ -17534,13 +18001,13 @@ async function evaluate(config) {
   let evalCases;
   let testFilePath;
   if (config.specFile) {
-    testFilePath = import_node_path42.default.resolve(config.specFile);
+    testFilePath = import_node_path43.default.resolve(config.specFile);
     evalCases = await loadTests(testFilePath, repoRoot, {
       verbose: config.verbose,
       filter: config.filter
     });
   } else {
-    testFilePath = import_node_path42.default.join(process.cwd(), "__programmatic__.yaml");
+    testFilePath = import_node_path43.default.join(process.cwd(), "__programmatic__.yaml");
     evalCases = (config.tests ?? []).map((test) => {
       const input = typeof test.input === "string" ? [{ role: "user", content: test.input }] : test.input;
       const question = typeof test.input === "string" ? test.input : test.input.find((m) => m.role === "user")?.content ?? "";
@@ -17626,10 +18093,10 @@ function computeSummary(results, durationMs) {
 var TARGET_FILE_CANDIDATES = [".agentv/targets.yaml", ".agentv/targets.yml"];
 async function discoverDefaultTarget(repoRoot) {
   const cwd = process.cwd();
-  const chain = buildDirectoryChain2(import_node_path42.default.join(cwd, "_placeholder"), repoRoot);
+  const chain = buildDirectoryChain2(import_node_path43.default.join(cwd, "_placeholder"), repoRoot);
   for (const dir of chain) {
     for (const candidate of TARGET_FILE_CANDIDATES) {
-      const targetsPath = import_node_path42.default.join(dir, candidate);
+      const targetsPath = import_node_path43.default.join(dir, candidate);
       if (!(0, import_node_fs12.existsSync)(targetsPath)) continue;
       try {
         const definitions = await readTargetDefinitions(targetsPath);
@@ -17644,10 +18111,10 @@ async function discoverDefaultTarget(repoRoot) {
 async function loadEnvHierarchy(repoRoot) {
   const { readFileSync: readFileSync2 } = await import("fs");
   const cwd = process.cwd();
-  const chain = buildDirectoryChain2(import_node_path42.default.join(cwd, "_placeholder"), repoRoot);
+  const chain = buildDirectoryChain2(import_node_path43.default.join(cwd, "_placeholder"), repoRoot);
   const envFiles = [];
   for (const dir of chain) {
-    const envPath = import_node_path42.default.join(dir, ".env");
+    const envPath = import_node_path43.default.join(dir, ".env");
     if ((0, import_node_fs12.existsSync)(envPath)) envFiles.push(envPath);
   }
   for (let i = envFiles.length - 1; i >= 0; i--) {
@@ -17829,7 +18296,7 @@ function buildPrompt(criteria, question, referenceAnswer) {
 
 // src/evaluation/cache/response-cache.ts
 var import_promises30 = require("fs/promises");
-var import_node_path43 = __toESM(require("path"), 1);
+var import_node_path44 = __toESM(require("path"), 1);
 var DEFAULT_CACHE_PATH = ".agentv/cache";
 var ResponseCache = class {
   cachePath;
@@ -17847,13 +18314,13 @@ var ResponseCache = class {
   }
   async set(key, value) {
     const filePath = this.keyToPath(key);
-    const dir = import_node_path43.default.dirname(filePath);
+    const dir = import_node_path44.default.dirname(filePath);
     await (0, import_promises30.mkdir)(dir, { recursive: true });
     await (0, import_promises30.writeFile)(filePath, JSON.stringify(value, null, 2), "utf8");
   }
   keyToPath(key) {
     const prefix = key.slice(0, 2);
-    return import_node_path43.default.join(this.cachePath, prefix, `${key}.json`);
+    return import_node_path44.default.join(this.cachePath, prefix, `${key}.json`);
   }
 };
 function shouldEnableCache(params) {
@@ -18340,6 +18807,7 @@ function createAgentKernel() {
   TokenUsageEvaluator,
   ToolTrajectoryEvaluator,
   WorkspaceCreationError,
+  WorkspacePoolManager,
   assembleLlmJudgePrompt,
   avgToolDurationMs,
   buildDirectoryChain,
@@ -18354,6 +18822,7 @@ function createAgentKernel() {
   cleanupEvalWorkspaces,
   cleanupWorkspace,
   computeTraceSummary,
+  computeWorkspaceFingerprint,
   consumeClaudeLogEntries,
   consumeCodexLogEntries,
   consumeCopilotCliLogEntries,
@@ -18386,11 +18855,11 @@ function createAgentKernel() {
   freeformEvaluationSchema,
   generateRubrics,
   getAgentvHome,
-  getGitCacheRoot,
   getHitCount,
   getSubagentsRoot,
   getTraceStateRoot,
   getWorkspacePath,
+  getWorkspacePoolRoot,
   getWorkspacesRoot,
   initializeBaseline,
   isEvaluatorKind,