@brainbase-labs/cli 0.24.0 → 0.25.0-eng1209.2

package/dist/index.js

@@ -36008,7 +36008,7 @@ function padStart(s, n) {
 // package.json
 var package_default = {
   name: "@brainbase-labs/cli",
-  version: "0.24.0",
+  version: "0.25.0-eng1209.2",
   description: "Pack, share, and install agent templates across harnesses (Claude Code, Codex, ...).",
   type: "module",
   bin: {
@@ -78485,6 +78485,13 @@ var MAX_TRAJECTORY_BYTES = 100 * 1024 * 1024;
 var MAX_REMOTE_INPUT_BYTES = 2 * 1024 * 1024 * 1024;
 var MAX_ARCHIVE_EXTRACTED_BYTES = 2 * 1024 * 1024 * 1024;
 var MAX_ARCHIVE_SCAN_BYTES = 2 * 1024 * 1024 * 1024;
+var MAX_SANDBOX_COMMANDS = 20;
+var MAX_CRITERIA_PER_EVALUATOR = 100;
+var MAX_CRITERIA_RESULT_BYTES = 1024 * 1024;
+var MAX_CRITERION_EXPLANATION_LENGTH = 16384;
+var MAX_CRITERION_EVIDENCE_IDS = 100;
+var MAX_CRITERION_EVIDENCE_ID_LENGTH = 256;
+var MAX_ALLOWED_EVIDENCE_IDS = 1e4;
 var RESERVED_WORKSPACE_PATHS = new Set([
   ".brainbase",
   ".git",
@@ -78512,6 +78519,17 @@ var NonSecretEnvironmentValueSchema = exports_external.object({
   sensitive: exports_external.literal(false)
 }).strict();
 var RootRelativePathSchema = exports_external.string().min(1).max(1024);
+var CandidateOutputPatternSchema = RootRelativePathSchema.refine((value) => {
+  if (value.includes("\\") || /[\[\]{}]/.test(value))
+    return false;
+  try {
+    return safeRelPath(value) === value;
+  } catch {
+    return false;
+  }
+}, {
+  message: "must be a normalized safe relative glob using only *, ?, and ** wildcards"
+});
 var ArchivePathSchema = RootRelativePathSchema.refine((value) => {
   try {
     return safeRelPath(value) === value;
@@ -78618,7 +78636,11 @@ var EvaluatorSchema = exports_external.discriminatedUnion("type", [
     ...EvaluatorBase,
     type: exports_external.literal("sandbox_command"),
     command: CommandSchema.omit({ id: true }),
-    root: exports_external.enum(["workspace", "tests"]).default("tests")
+    root: exports_external.enum(["workspace", "tests"]).default("tests"),
+    criterion_keys: exports_external.array(IdSchema).min(1).max(MAX_CRITERIA_PER_EVALUATOR).optional(),
+    allowed_evidence_ids: exports_external.array(exports_external.string().min(1).max(MAX_CRITERION_EVIDENCE_ID_LENGTH)).max(MAX_ALLOWED_EVIDENCE_IDS).optional(),
+    tests_path: ArchivePathSchema.optional(),
+    workspace_mode: exports_external.enum(["read_only", "isolated_copy"]).default("read_only")
   }).strict()
 ]);
 var EvidenceFileSchema = exports_external.object({
@@ -78638,6 +78660,37 @@ var HydrateSpecSchema = BaseSpecSchema.extend({
   materials: exports_external.array(MaterialSchema).max(1e4).default([]),
   setup_commands: exports_external.array(CommandSchema).max(128).default([])
 }).strict();
+var CandidateOutputSchema = exports_external.object({
+  id: IdSchema,
+  pattern: CandidateOutputPatternSchema,
+  kind: exports_external.enum(["file", "directory"]),
+  required: exports_external.boolean(),
+  min_matches: exports_external.number().int().min(0).max(1e5),
+  max_matches: exports_external.number().int().min(1).max(1e5),
+  max_total_bytes: exports_external.number().int().min(0).max(2 * 1024 * 1024 * 1024)
+}).strict().superRefine((value, context) => {
+  if (value.min_matches > value.max_matches) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["min_matches"],
+      message: "min_matches cannot exceed max_matches"
+    });
+  }
+  if (value.required && value.min_matches === 0) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["min_matches"],
+      message: "required candidate outputs must require at least one match"
+    });
+  }
+  if (!value.required && value.min_matches > 0) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["min_matches"],
+      message: "optional candidate outputs must allow zero matches"
+    });
+  }
+});
 var EvaluateSpecSchema = BaseSpecSchema.extend({
   phase: exports_external.literal("evaluate"),
   tests_root: AbsolutePathSchema,
@@ -78648,6 +78701,7 @@ var EvaluateSpecSchema = BaseSpecSchema.extend({
   references: exports_external.array(MaterialSchema).max(1e4).default([]),
   evaluators: exports_external.array(EvaluatorSchema).min(1).max(1000),
   candidate_artifacts: exports_external.array(RootRelativePathSchema).max(1000).default([]),
+  candidate_outputs: exports_external.array(CandidateOutputSchema).max(1000).default([]),
   capture_workspace_archive: exports_external.boolean().default(false),
   workspace_limits: exports_external.object({
     max_file_count: exports_external.number().int().min(1).max(1e6).default(1e5),
@@ -78727,7 +78781,8 @@ var BenchmarkSpecSchema = exports_external.discriminatedUnion("phase", [
     { items: value.setup_commands, path: "setup_commands" }
   ] : [
     { items: value.references, path: "references" },
-    { items: value.evaluators, path: "evaluators" }
+    { items: value.evaluators, path: "evaluators" },
+    { items: value.candidate_outputs, path: "candidate_outputs" }
   ];
   for (const { items, path: issuePath } of uniqueLists) {
     const ids = items.map((item) => item.id);
@@ -78739,6 +78794,39 @@ var BenchmarkSpecSchema = exports_external.discriminatedUnion("phase", [
       });
     }
   }
+  if (value.phase === "evaluate") {
+    value.evaluators.forEach((evaluator, index) => {
+      if (evaluator.type === "sandbox_command" && evaluator.criterion_keys && new Set(evaluator.criterion_keys).size !== evaluator.criterion_keys.length) {
+        context.addIssue({
+          code: exports_external.ZodIssueCode.custom,
+          path: ["evaluators", index, "criterion_keys"],
+          message: "criterion_keys must be unique"
+        });
+      }
+      if (evaluator.type === "sandbox_command" && evaluator.criterion_keys && evaluator.allowed_evidence_ids === undefined) {
+        context.addIssue({
+          code: exports_external.ZodIssueCode.custom,
+          path: ["evaluators", index, "allowed_evidence_ids"],
+          message: "structured criteria require allowed_evidence_ids"
+        });
+      }
+      if (evaluator.type === "sandbox_command" && evaluator.allowed_evidence_ids && new Set(evaluator.allowed_evidence_ids).size !== evaluator.allowed_evidence_ids.length) {
+        context.addIssue({
+          code: exports_external.ZodIssueCode.custom,
+          path: ["evaluators", index, "allowed_evidence_ids"],
+          message: "allowed_evidence_ids must be unique"
+        });
+      }
+    });
+    const candidateOutputBytes = value.candidate_outputs.reduce((total, output) => total + output.max_total_bytes, 0);
+    if (candidateOutputBytes > value.workspace_limits.max_total_bytes) {
+      context.addIssue({
+        code: exports_external.ZodIssueCode.custom,
+        path: ["candidate_outputs"],
+        message: "candidate output byte limits exceed the workspace byte limit"
+      });
+    }
+  }
   const materials = value.phase === "hydrate" ? value.materials : value.references;
   const archiveOutputBytes = materials.reduce((total, material) => total + (material.kind === "archive_file" ? material.file_size_bytes : 0), 0);
   if (archiveOutputBytes > MAX_ARCHIVE_EXTRACTED_BYTES) {
@@ -78748,11 +78836,11 @@ var BenchmarkSpecSchema = exports_external.discriminatedUnion("phase", [
       message: "archive-file materials exceed the aggregate extracted byte limit"
     });
   }
-  if (value.phase === "evaluate" && value.evaluators.filter((evaluator) => evaluator.type === "sandbox_command").length > 1) {
+  if (value.phase === "evaluate" && value.evaluators.filter((evaluator) => evaluator.type === "sandbox_command").length > MAX_SANDBOX_COMMANDS) {
     context.addIssue({
       code: exports_external.ZodIssueCode.custom,
       path: ["evaluators"],
-      message: "schema version 1 supports at most one sandbox_command evaluator"
+      message: `schema version 1 supports at most ${MAX_SANDBOX_COMMANDS} sandbox_command evaluators`
     });
   }
 });
@@ -78763,7 +78851,12 @@ var BENCHMARK_CAPABILITIES = {
   phases: ["hydrate", "evaluate"],
   features: [
     "remote_input_references_v1",
-    "archive_file_materials_v1"
+    "archive_file_materials_v1",
+    "structured_criterion_results_v1",
+    "multiple_sandbox_commands_v1",
+    "sandbox_command_workspace_modes_v1",
+    "candidate_outputs_v1",
+    "criterion_evidence_allowlist_v1"
   ],
   evaluator_types: [
     "output_assertion",
@@ -78774,7 +78867,8 @@ var BENCHMARK_CAPABILITIES = {
   limits: {
     max_secret_bindings: 100,
     max_evaluators: 1000,
-    max_sandbox_commands: 1
+    max_sandbox_commands: MAX_SANDBOX_COMMANDS,
+    max_criteria_per_evaluator: MAX_CRITERIA_PER_EVALUATOR
   }
 };
 
@@ -80126,6 +80220,234 @@ async function workspaceManifest(spec, context) {
   }
   return records.sort((a3, b4) => a3.path.localeCompare(b4.path));
 }
+function candidateGlob(pattern) {
+  let source = "^";
+  for (let index = 0;index < pattern.length; index += 1) {
+    const char = pattern[index];
+    if (char === "*") {
+      if (pattern[index + 1] === "*") {
+        if (pattern[index + 2] === "/") {
+          source += "(?:.*/)?";
+          index += 2;
+        } else {
+          source += ".*";
+          index += 1;
+        }
+      } else {
+        source += "[^/]*";
+      }
+      continue;
+    }
+    if (char === "?") {
+      source += "[^/]";
+      continue;
+    }
+    source += char.replace(/[\\^$.*+?()[\]{}|]/g, "\\$&");
+  }
+  return new RegExp(`${source}$`);
+}
+function manifestDirectories(manifest, context) {
+  const directories = new Set;
+  for (const entry of manifest) {
+    assertBudget(context);
+    let current = path88.posix.dirname(entry.path);
+    while (current !== ".") {
+      assertBudget(context);
+      directories.add(current);
+      current = path88.posix.dirname(current);
+    }
+  }
+  return [...directories].sort();
+}
+function candidateOutputFiles(output, manifest, context) {
+  const matcher = candidateGlob(output.pattern);
+  if (output.kind === "file") {
+    const matched = [];
+    for (const entry of manifest) {
+      assertBudget(context);
+      if (!matcher.test(entry.path))
+        continue;
+      if (entry.kind === "symlink") {
+        throw new BenchmarkPhaseError("unsafe_path", `candidate output ${output.id} matches a symlink`);
+      }
+      matched.push(entry);
+    }
+    return {
+      matchedCount: matched.length,
+      files: matched
+    };
+  }
+  for (const entry of manifest) {
+    assertBudget(context);
+    if (entry.kind === "symlink" && matcher.test(entry.path)) {
+      throw new BenchmarkPhaseError("unsafe_path", `candidate output ${output.id} matches a symlink`);
+    }
+  }
+  const directories = [];
+  for (const directory of manifestDirectories(manifest, context)) {
+    assertBudget(context);
+    if (matcher.test(directory))
+      directories.push(directory);
+  }
+  const selected = new Map;
+  for (const directory of directories) {
+    assertBudget(context);
+    const prefix = `${directory}/`;
+    for (const entry of manifest) {
+      assertBudget(context);
+      if (!entry.path.startsWith(prefix))
+        continue;
+      if (entry.kind === "symlink") {
+        throw new BenchmarkPhaseError("unsafe_path", `candidate output ${output.id} contains a symlink`);
+      }
+      selected.set(entry.path, entry);
+    }
+  }
+  return {
+    matchedCount: directories.length,
+    files: [...selected.values()].sort((left, right) => left.path.localeCompare(right.path))
+  };
+}
+async function copyCandidateOutput(output, manifest, spec, context) {
+  const selected = candidateOutputFiles(output, manifest, context);
+  if (!output.required && selected.matchedCount === 0)
+    return [];
+  const minimumMatches = output.required ? Math.max(1, output.min_matches) : output.min_matches;
+  if (selected.matchedCount < minimumMatches || selected.matchedCount > output.max_matches) {
+    throw new BenchmarkPhaseError("candidate_output_match_count", `candidate output ${output.id} matched ${selected.matchedCount} ${output.kind}s`, {
+      actual: selected.matchedCount,
+      min_matches: minimumMatches,
+      max_matches: output.max_matches
+    });
+  }
+  let totalBytes = 0;
+  for (const file of selected.files) {
+    assertBudget(context);
+    totalBytes += file.size;
+  }
+  if (totalBytes > output.max_total_bytes) {
+    throw new BenchmarkPhaseError("candidate_output_size_exceeded", `candidate output ${output.id} exceeds its byte limit`, { actual: totalBytes, max_total_bytes: output.max_total_bytes });
+  }
+  const copied = [];
+  for (const frozenFile of selected.files) {
+    assertBudget(context);
+    const source = path88.resolve(spec.workspace_root, safeRelPath(frozenFile.path));
+    const destination = path88.resolve(spec.logs_root, "candidate-outputs", output.id, safeRelPath(frozenFile.path));
+    if (fs81.existsSync(destination)) {
+      throw new BenchmarkPhaseError("destination_conflict", `candidate output destination already exists: ${output.id}/${frozenFile.path}`);
+    }
+    await atomicCopy(source, destination, frozenFile.mode, spec.workspace_root);
+    const record3 = await recordFile(spec.logs_root, destination, "logs");
+    if (record3.sha256 !== frozenFile.sha256 || record3.size !== frozenFile.size || record3.mode !== frozenFile.mode) {
+      throw new BenchmarkPhaseError("evidence_tampered", `candidate output changed after the workspace freeze: ${frozenFile.path}`);
+    }
+    copied.push(record3);
+  }
+  return copied;
+}
+async function copyFrozenWorkspace(manifest, spec, context) {
+  for (const entry of manifest) {
+    assertBudget(context);
+    if (entry.kind === "symlink") {
+      throw new BenchmarkPhaseError("unsupported_workspace_symlink", `sandbox evaluator workspace cannot safely reproduce symlink: ${entry.path}`);
+    }
+  }
+  const destinationRoot = fs81.mkdtempSync(path88.join(os16.tmpdir(), "brainbase-benchmark-workspace-"));
+  fs81.chmodSync(destinationRoot, 448);
+  context.temporaryRoots.add(destinationRoot);
+  for (const frozenFile of manifest) {
+    assertBudget(context);
+    const source = path88.resolve(spec.workspace_root, safeRelPath(frozenFile.path));
+    const destination = path88.resolve(destinationRoot, safeRelPath(frozenFile.path));
+    await atomicCopy(source, destination, frozenFile.mode, spec.workspace_root);
+    const copied = await recordFile(destinationRoot, destination, "workspace");
+    if (copied.sha256 !== frozenFile.sha256 || copied.size !== frozenFile.size || copied.mode !== frozenFile.mode) {
+      throw new BenchmarkPhaseError("evidence_tampered", `workspace changed after it was frozen: ${frozenFile.path}`);
+    }
+  }
+  return destinationRoot;
+}
+function evaluatorTestsPath(evaluator, spec) {
+  if (!evaluator.tests_path)
+    return spec.tests_root;
+  const relative = normalizedRootRelative(evaluator.tests_path);
+  assertNoSymlinkTraversal(spec.tests_root, relative);
+  const candidate = path88.resolve(spec.tests_root, relative);
+  if (!isWithin(spec.tests_root, candidate) || !fs81.existsSync(candidate)) {
+    throw new BenchmarkPhaseError("invalid_tests_path", `evaluator tests_path does not exist: ${evaluator.tests_path}`);
+  }
+  const stat = fs81.lstatSync(candidate);
+  if (stat.isSymbolicLink() || !stat.isDirectory()) {
+    throw new BenchmarkPhaseError("invalid_tests_path", `evaluator tests_path must be a directory: ${evaluator.tests_path}`);
+  }
+  return candidate;
+}
+var CriterionResultSchema = exports_external.object({
+  criterion_key: IdSchema,
+  outcome: exports_external.enum(["satisfied", "not_satisfied", "not_applicable"]),
+  explanation: exports_external.string().min(1).max(MAX_CRITERION_EXPLANATION_LENGTH),
+  evidence_ids: exports_external.array(exports_external.string().min(1).max(MAX_CRITERION_EVIDENCE_ID_LENGTH)).max(MAX_CRITERION_EVIDENCE_IDS)
+}).strict().superRefine((value, context) => {
+  if (new Set(value.evidence_ids).size !== value.evidence_ids.length) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["evidence_ids"],
+      message: "evidence_ids must be unique"
+    });
+  }
+});
+var CriteriaResultFileSchema = exports_external.object({
+  criteria: exports_external.array(CriterionResultSchema).max(MAX_CRITERIA_PER_EVALUATOR)
+}).strict();
+function readCriterionResults(resultPath, criterionKeys, allowedEvidenceIds) {
+  let opened;
+  try {
+    opened = openRegularFileNoFollow(resultPath, "criterion result", path88.dirname(resultPath));
+  } catch (error2) {
+    if (error2.code === "ENOENT") {
+      throw new BenchmarkPhaseError("missing_criterion_result", "sandbox evaluator did not write its criterion result");
+    }
+    throw error2;
+  }
+  try {
+    if (opened.stat.size > MAX_CRITERIA_RESULT_BYTES) {
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result exceeds the 1 MiB limit");
+    }
+    let raw;
+    try {
+      raw = JSON.parse(readDescriptor(opened.fd).toString("utf8"));
+    } catch {
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result is not valid JSON");
+    }
+    let parsed;
+    try {
+      parsed = CriteriaResultFileSchema.parse(raw);
+    } catch (error2) {
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result does not match the required schema", error2 instanceof exports_external.ZodError ? error2.issues : undefined);
+    }
+    const byKey = new Map;
+    for (const criterion of parsed.criteria) {
+      if (byKey.has(criterion.criterion_key)) {
+        throw new BenchmarkPhaseError("invalid_criterion_result", `criterion result is duplicated: ${criterion.criterion_key}`);
+      }
+      byKey.set(criterion.criterion_key, criterion);
+    }
+    const expected = new Set(criterionKeys);
+    const extra = [...byKey.keys()].filter((key2) => !expected.has(key2));
+    const missing = criterionKeys.filter((key2) => !byKey.has(key2));
+    if (extra.length > 0 || missing.length > 0) {
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result keys do not match the evaluator plan", { missing, extra });
+    }
+    const allowedEvidence = new Set(allowedEvidenceIds);
+    const unknownEvidenceIds = [...new Set(parsed.criteria.flatMap((criterion) => criterion.evidence_ids))].filter((evidenceId) => !allowedEvidence.has(evidenceId));
+    if (unknownEvidenceIds.length > 0) {
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result cites evidence outside the evaluator plan", { unknown_evidence_ids: unknownEvidenceIds });
+    }
+    return criterionKeys.map((key2) => byKey.get(key2));
+  } finally {
+    fs81.closeSync(opened.fd);
+  }
+}
 function trajectoryEvents(value) {
   if (Array.isArray(value))
     return value;
@@ -80144,7 +80466,7 @@ function eventType(event) {
   }
   return;
 }
-async function evaluateOne(evaluator, spec, finalOutput, trajectory, frozenEvidence, context) {
+async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, frozenEvidence, context) {
   const started = Date.now();
   const base2 = {
     id: evaluator.id,
@@ -80156,11 +80478,11 @@ async function evaluateOne(evaluator, spec, finalOutput, trajectory, frozenEvide
   };
   if (evaluator.type === "output_assertion") {
     const assertion = evaluator.assertion;
-    let verdict2 = false;
+    let verdict = false;
     if (assertion.operator === "exact")
-      verdict2 = finalOutput === assertion.expected;
+      verdict = finalOutput === assertion.expected;
     if (assertion.operator === "contains")
-      verdict2 = finalOutput.includes(assertion.expected);
+      verdict = finalOutput.includes(assertion.expected);
     if (assertion.operator === "regex") {
       const regexResult = await runCommand({
         id: `${evaluator.id}.regex`,
@@ -80186,17 +80508,17 @@ async function evaluateOne(evaluator, spec, finalOutput, trajectory, frozenEvide
       if (regexResult.exitCode === 2) {
         throw new BenchmarkPhaseError("invalid_evaluator", `invalid regex in evaluator: ${evaluator.id}`);
       }
-      verdict2 = regexResult.exitCode === 0;
+      verdict = regexResult.exitCode === 0;
     }
-    return { ...base2, status: verdict2 ? "passed" : "failed", verdict: verdict2, duration_ms: Date.now() - started };
+    return { ...base2, status: verdict ? "passed" : "failed", verdict, duration_ms: Date.now() - started };
   }
   if (evaluator.type === "trajectory_assertion") {
     const count = trajectory.filter((event) => eventType(event) === evaluator.event_type).length;
-    const verdict2 = count >= evaluator.min_count && (evaluator.max_count === undefined || count <= evaluator.max_count);
+    const verdict = count >= evaluator.min_count && (evaluator.max_count === undefined || count <= evaluator.max_count);
     return {
       ...base2,
-      status: verdict2 ? "passed" : "failed",
-      verdict: verdict2,
+      status: verdict ? "passed" : "failed",
+      verdict,
       duration_ms: Date.now() - started,
       details: { count, min_count: evaluator.min_count, max_count: evaluator.max_count }
     };
@@ -80217,16 +80539,16 @@ async function evaluateOne(evaluator, spec, finalOutput, trajectory, frozenEvide
       throw new BenchmarkPhaseError("unsafe_path", `workspace assertion cannot target a symlink: ${relative}`);
     }
     const exists2 = stat !== null;
-    let verdict2 = false;
+    let verdict = false;
     if (evaluator.assertion.operator === "exists")
-      verdict2 = exists2;
+      verdict = exists2;
     if (evaluator.assertion.operator === "not_exists")
-      verdict2 = !exists2;
+      verdict = !exists2;
     if (evaluator.assertion.operator === "sha256") {
       if (stat?.isFile()) {
         const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, spec.workspace_root);
         try {
-          verdict2 = await sha256OfDescriptor(opened.fd) === evaluator.assertion.expected;
+          verdict = await sha256OfDescriptor(opened.fd) === evaluator.assertion.expected;
         } finally {
           fs81.closeSync(opened.fd);
         }
@@ -80236,35 +80558,78 @@ async function evaluateOne(evaluator, spec, finalOutput, trajectory, frozenEvide
       if (stat?.isFile()) {
         const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, spec.workspace_root);
         try {
-          verdict2 = readDescriptor(opened.fd).toString("utf8").includes(evaluator.assertion.expected);
+          verdict = readDescriptor(opened.fd).toString("utf8").includes(evaluator.assertion.expected);
         } finally {
           fs81.closeSync(opened.fd);
         }
       }
     }
-    return { ...base2, status: verdict2 ? "passed" : "failed", verdict: verdict2, duration_ms: Date.now() - started };
+    return { ...base2, status: verdict ? "passed" : "failed", verdict, duration_ms: Date.now() - started };
+  }
+  let isolatedWorkspace;
+  let privateResultRoot;
+  try {
+    const evaluatorWorkspace = await copyFrozenWorkspace(manifest, spec, context);
+    isolatedWorkspace = evaluatorWorkspace;
+    const testsPath = evaluatorTestsPath(evaluator, spec);
+    const commandRoot = evaluator.root === "workspace" ? evaluatorWorkspace : spec.tests_root;
+    const command = { id: evaluator.id, ...evaluator.command };
+    const environment = {
+      BRAINBASE_BENCHMARK_WORKSPACE: evaluatorWorkspace,
+      BRAINBASE_BENCHMARK_TESTS: testsPath,
+      BRAINBASE_BENCHMARK_LOGS: spec.logs_root,
+      BRAINBASE_BENCHMARK_FINAL_OUTPUT: frozenEvidence.finalOutputPath,
+      BRAINBASE_BENCHMARK_TRAJECTORY: frozenEvidence.trajectoryPath
+    };
+    let criterionResultPath;
+    if (evaluator.criterion_keys) {
+      privateResultRoot = fs81.mkdtempSync(path88.join(os16.tmpdir(), "brainbase-benchmark-criteria-"));
+      fs81.chmodSync(privateResultRoot, 448);
+      context.temporaryRoots.add(privateResultRoot);
+      criterionResultPath = path88.join(privateResultRoot, "result.json");
+      environment.BRAINBASE_BENCHMARK_CRITERIA_RESULT = criterionResultPath;
+    }
+    const result2 = await runCommand(command, commandRoot, spec, context, environment);
+    const stdout = await writeLog(spec.logs_root, `${evaluator.id}.stdout.log`, result2.stdout, spec);
+    const stderr = await writeLog(spec.logs_root, `${evaluator.id}.stderr.log`, result2.stderr, spec);
+    let criterionResults;
+    try {
+      criterionResults = evaluator.criterion_keys && criterionResultPath ? readCriterionResults(criterionResultPath, evaluator.criterion_keys, evaluator.allowed_evidence_ids ?? []) : undefined;
+    } catch (error2) {
+      const normalized = stableError(error2);
+      return {
+        ...base2,
+        status: "errored",
+        verdict: null,
+        duration_ms: result2.durationMs,
+        details: {
+          exit_code: result2.exitCode,
+          error_code: normalized?.code ?? "invalid_criterion_result",
+          error_message: normalized?.message ?? "criterion result validation failed"
+        },
+        stdout,
+        stderr
+      };
+    }
+    const verdict = result2.exitCode === 0;
+    return {
+      ...base2,
+      status: verdict ? "passed" : "failed",
+      verdict,
+      duration_ms: result2.durationMs,
+      details: { exit_code: result2.exitCode },
+      stdout,
+      stderr,
+      ...criterionResults ? { criterion_results: criterionResults } : {}
+    };
+  } finally {
+    for (const temporary of [privateResultRoot, isolatedWorkspace]) {
+      if (!temporary)
+        continue;
+      fs81.rmSync(temporary, { recursive: true, force: true });
+      context.temporaryRoots.delete(temporary);
+    }
   }
-  const commandRoot = evaluator.root === "workspace" ? spec.workspace_root : spec.tests_root;
-  const command = { id: evaluator.id, ...evaluator.command };
-  const result2 = await runCommand(command, commandRoot, spec, context, {
-    BRAINBASE_BENCHMARK_WORKSPACE: spec.workspace_root,
-    BRAINBASE_BENCHMARK_TESTS: spec.tests_root,
-    BRAINBASE_BENCHMARK_LOGS: spec.logs_root,
-    BRAINBASE_BENCHMARK_FINAL_OUTPUT: frozenEvidence.finalOutputPath,
-    BRAINBASE_BENCHMARK_TRAJECTORY: frozenEvidence.trajectoryPath
-  });
-  const stdout = await writeLog(spec.logs_root, `${evaluator.id}.stdout.log`, result2.stdout, spec);
-  const stderr = await writeLog(spec.logs_root, `${evaluator.id}.stderr.log`, result2.stderr, spec);
-  const verdict = result2.exitCode === 0;
-  return {
-    ...base2,
-    status: verdict ? "passed" : "failed",
-    verdict,
-    duration_ms: result2.durationMs,
-    details: { exit_code: result2.exitCode },
-    stdout,
-    stderr
-  };
 }
 async function executeEvaluate(spec, context) {
   await downloadInputReference(spec.staging_root, spec.evidence.final_output, context);
@@ -80336,6 +80701,11 @@ async function executeEvaluate(spec, context) {
     outputs.push(artifact);
     context.outputs.push(artifact);
   }
+  for (const candidateOutput of spec.candidate_outputs) {
+    const copied = await copyCandidateOutput(candidateOutput, manifest, spec, context);
+    outputs.push(...copied);
+    context.outputs.push(...copied);
+  }
   if (spec.capture_workspace_archive) {
     const regularFiles = manifest.filter((entry) => entry.kind !== "symlink").map((entry) => entry.path);
     const archive = path88.join(spec.logs_root, "candidate-workspace.tar.gz");
@@ -80363,8 +80733,9 @@ async function executeEvaluate(spec, context) {
   for (const evaluator of executionOrder) {
     assertBudget(context);
     const started = Date.now();
+    let requiredResultError;
     try {
-      const evaluated = await evaluateOne(evaluator, spec, finalOutput.buffer.toString("utf8"), trajectory, { finalOutputPath, trajectoryPath }, context);
+      const evaluated = await evaluateOne(evaluator, spec, manifest, finalOutput.buffer.toString("utf8"), trajectory, { finalOutputPath, trajectoryPath }, context);
       assertBudget(context);
       evaluators.push(evaluated);
       context.evaluators.push(evaluated);
@@ -80376,6 +80747,11 @@ async function executeEvaluate(spec, context) {
         outputs.push(evaluated.stderr);
         context.outputs.push(evaluated.stderr);
       }
+      if (evaluated.status === "errored" && evaluator.required) {
+        const errorCode = typeof evaluated.details?.error_code === "string" ? evaluated.details.error_code : "evaluator_execution_failed";
+        const errorMessage2 = typeof evaluated.details?.error_message === "string" ? evaluated.details.error_message : "required evaluator execution failed";
+        requiredResultError = new BenchmarkPhaseError(errorCode, errorMessage2);
+      }
     } catch (error2) {
       const normalized = stableError(error2);
       const errored = {
@@ -80399,6 +80775,8 @@ async function executeEvaluate(spec, context) {
         throw error2;
       assertBudget(context);
     }
+    if (requiredResultError)
+      throw requiredResultError;
   }
   await verifyRecordsUnchanged([...context.inputs, ...context.outputs.slice(0, frozenOutputCount)], spec);
   if (!verifyOwnedDirectory(spec.tests_root, "tests", spec) || !verifyOwnedDirectory(spec.logs_root, "logs", spec)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@brainbase-labs/cli",
-  "version": "0.24.0",
+  "version": "0.25.0-eng1209.2",
   "description": "Pack, share, and install agent templates across harnesses (Claude Code, Codex, ...).",
   "type": "module",
   "bin": {