@brainbase-labs/cli 0.25.0-eng1209.1 → 0.25.0-eng1209.11

package/README.md

@@ -72,21 +72,28 @@ synchronously. Schema v1 supports:
 - `workspace_assertion`: file existence, absence, SHA-256, or content checks.
 - `sandbox_command`: a bounded argv command rooted in the workspace or hidden
   tests directory. Exit zero passes; any other exit code is a valid failed
-  verdict. Schema v1 allows one command evaluator and runs it after all
-  read-only assertions so evaluator mutations cannot change candidate verdicts.
+  verdict. Schema v1 allows up to 20 command evaluators and runs them after all
+  read-only assertions. Each command receives either the verified live workspace
+  for legacy read-only evaluation or its own isolated frozen workspace copy.
 
 Valid failed verdicts still produce a successful evaluation phase. Invalid
-specs, digest/path violations, missing environment, launch failures, timeouts,
-and output-budget violations fail the phase. Results are written atomically
+specs, digest/path violations, missing environment, phase-budget exhaustion,
+and output-budget violations fail the phase. Individual command launch failures,
+timeouts, and forced terminations are recorded as errored evaluator results so
+the remaining evaluators can still run. Results are written atomically
 with mode `0600`, include the raw spec digest and checksummed evidence/log
 input/output references with explicit `staging`, `workspace`, `tests`, or
 `logs` roots, redact declared secret values from command logs, require the
 result path to be `<logs_root>/result.json`, and treat a matching successful
 phase result as authoritative on replay.
 
-On any phase-level execution failure, MAS must tear down the task sandbox. The
-CLI kills the command process group and descendants it can observe, but the
-runtime lifecycle remains the final cleanup boundary for daemonized processes.
+On any phase-level execution failure, MAS must tear down the task sandbox.
+Failed and timed-out commands clean up the process group and descendants the
+CLI can observe. Successful evaluator commands receive the same cleanup so one
+check cannot contaminate the next. Successful hydration setup commands may
+intentionally leave services running for the agent turn; the runtime lifecycle
+remains their final cleanup boundary. Isolated evaluator workspace copies omit
+`.git` and unsafe absolute or escaping symlinks.
 
 ## Development
 

package/dist/index.js

@@ -36008,7 +36008,7 @@ function padStart(s, n) {
 // package.json
 var package_default = {
   name: "@brainbase-labs/cli",
-  version: "0.25.0-eng1209.1",
+  version: "0.25.0-eng1209.11",
   description: "Pack, share, and install agent templates across harnesses (Claude Code, Codex, ...).",
   type: "module",
   bin: {
@@ -78485,13 +78485,20 @@ 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_HYDRATE_COMMANDS = 150;
+var MAX_PHASE_RESULT_BYTES = 32 * 1024 * 1024;
 var MAX_SANDBOX_COMMANDS = 20;
-var MAX_CRITERIA_PER_EVALUATOR = 100;
-var MAX_CRITERIA_RESULT_BYTES = 1024 * 1024;
+var MAX_CRITERIA_PER_EVALUATOR = 200;
+var MAX_CRITERIA_RESULT_BYTES = 24 * 1024 * 1024;
+var MAX_CRITERIA_RESULT_PAYLOAD_BYTES = 20 * 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 COMMAND_PROCESS_MARKER_ENV = "BRAINBASE_BENCHMARK_COMMAND_MARKER";
+var COMMAND_PROCESS_POLL_MS = 10;
+var COMMAND_PROCESS_CLEANUP_MS = 500;
+var JUDGE_ERROR_PREFIX = "BRAINBASE_BENCHMARK_JUDGE_ERROR_V1:";
 var RESERVED_WORKSPACE_PATHS = new Set([
   ".brainbase",
   ".git",
@@ -78658,7 +78665,7 @@ var TrajectoryEvidenceSchema = EvidenceFileSchema.extend({
 var HydrateSpecSchema = BaseSpecSchema.extend({
   phase: exports_external.literal("hydrate"),
   materials: exports_external.array(MaterialSchema).max(1e4).default([]),
-  setup_commands: exports_external.array(CommandSchema).max(128).default([])
+  setup_commands: exports_external.array(CommandSchema).max(MAX_HYDRATE_COMMANDS).default([])
 }).strict();
 var CandidateOutputSchema = exports_external.object({
   id: IdSchema,
@@ -78712,6 +78719,20 @@ var BenchmarkSpecSchema = exports_external.discriminatedUnion("phase", [
   HydrateSpecSchema,
   EvaluateSpecSchema
 ]).superRefine((value, context) => {
+  if (Object.hasOwn(value.environment, COMMAND_PROCESS_MARKER_ENV)) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["environment", COMMAND_PROCESS_MARKER_ENV],
+      message: `${COMMAND_PROCESS_MARKER_ENV} is reserved for command supervision`
+    });
+  }
+  if (value.secret_env.includes(COMMAND_PROCESS_MARKER_ENV)) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["secret_env"],
+      message: `${COMMAND_PROCESS_MARKER_ENV} is reserved for command supervision`
+    });
+  }
   for (const name of Object.keys(value.environment)) {
     if (SENSITIVE_ENV_NAME_RE.test(name)) {
       context.addIssue({
@@ -78785,12 +78806,12 @@ var BenchmarkSpecSchema = exports_external.discriminatedUnion("phase", [
     { items: value.candidate_outputs, path: "candidate_outputs" }
   ];
   for (const { items, path: issuePath } of uniqueLists) {
-    const ids = items.map((item) => item.id);
+    const ids = items.map((item) => issuePath === "candidate_outputs" ? item.id.toLowerCase() : item.id);
     if (new Set(ids).size !== ids.length) {
       context.addIssue({
         code: exports_external.ZodIssueCode.custom,
         path: [issuePath],
-        message: `${issuePath} ids must be unique`
+        message: `${issuePath} ids must be unique${issuePath === "candidate_outputs" ? " ignoring case" : ""}`
       });
     }
   }
@@ -78853,6 +78874,7 @@ var BENCHMARK_CAPABILITIES = {
     "remote_input_references_v1",
     "archive_file_materials_v1",
     "structured_criterion_results_v1",
+    "structured_judge_errors_v1",
     "multiple_sandbox_commands_v1",
     "sandbox_command_workspace_modes_v1",
     "candidate_outputs_v1",
@@ -78866,6 +78888,8 @@ var BENCHMARK_CAPABILITIES = {
   ],
   limits: {
     max_secret_bindings: 100,
+    max_hydrate_commands: MAX_HYDRATE_COMMANDS,
+    max_phase_result_bytes: MAX_PHASE_RESULT_BYTES,
     max_evaluators: 1000,
     max_sandbox_commands: MAX_SANDBOX_COMMANDS,
     max_criteria_per_evaluator: MAX_CRITERIA_PER_EVALUATOR
@@ -78882,6 +78906,19 @@ class BenchmarkPhaseError extends Error {
     this.name = "BenchmarkPhaseError";
   }
 }
+
+class BenchmarkCommandExecutionError extends BenchmarkPhaseError {
+  stdout;
+  stderr;
+  durationMs;
+  constructor(code, message, stdout, stderr, durationMs) {
+    super(code, message);
+    this.stdout = stdout;
+    this.stderr = stderr;
+    this.durationMs = durationMs;
+    this.name = "BenchmarkCommandExecutionError";
+  }
+}
 var ZIP_EOCD_SIGNATURE = 101010256;
 var ZIP_CENTRAL_SIGNATURE = 33639248;
 var ZIP_LOCAL_SIGNATURE = 67324752;
@@ -79114,10 +79151,15 @@ async function verifyRecordsUnchanged(records, spec) {
     const relative = safeRelPath(record3.path);
     assertNoSymlinkTraversal(root, relative);
     const candidate = path88.resolve(root, relative);
-    if (!isWithin(root, candidate) || !fs81.existsSync(candidate)) {
+    if (!isWithin(root, candidate)) {
+      throw new BenchmarkPhaseError("evidence_tampered", `evidence was removed during evaluation: ${record3.root}:${record3.path}`);
+    }
+    let stat;
+    try {
+      stat = fs81.lstatSync(candidate);
+    } catch {
       throw new BenchmarkPhaseError("evidence_tampered", `evidence was removed during evaluation: ${record3.root}:${record3.path}`);
     }
-    const stat = fs81.lstatSync(candidate);
     if (record3.kind === "symlink") {
       const target = stat.isSymbolicLink() ? fs81.readlinkSync(candidate) : null;
       if (target === null || Buffer.byteLength(target) !== record3.size || sha256(target) !== record3.sha256) {
@@ -79212,6 +79254,15 @@ function removeRemoteHydrationInputs(spec) {
     removedSources.add(material.source);
   }
 }
+function removePrivateHydrationSourcesBeforeSetup(spec, context) {
+  removeRemoteHydrationInputs(spec);
+  for (const temporary of context.temporaryRoots) {
+    fs81.rmSync(temporary, { recursive: true, force: true });
+  }
+  context.temporaryRoots.clear();
+  context.verifiedInputs.clear();
+  context.preparedArchiveFiles.clear();
+}
 async function downloadInputReference(stagingRoot, input, context) {
   const cacheKey = inputCacheKey(input);
   const cached2 = context.verifiedInputs.get(cacheKey);
@@ -79893,7 +79944,7 @@ function buildEnvironment(spec, secretNames, additions = {}) {
   Object.assign(env3, additions);
   return env3;
 }
-function redactCommandOutput(data, spec) {
+function redactCommandOutput(data, spec, additionalSecrets = []) {
   let value = data.toString("utf8");
   const sensitiveNames = new Set([
     ...spec.secret_env,
@@ -79904,6 +79955,10 @@ function redactCommandOutput(data, spec) {
     if (secret)
       value = value.split(secret).join("[REDACTED]");
   }
+  for (const secret of additionalSecrets) {
+    if (secret)
+      value = value.split(secret).join("[REDACTED]");
+  }
   return Buffer.from(value);
 }
 function descendantPids(parentPid) {
@@ -79955,7 +80010,99 @@ function terminate(child) {
     child.kill("SIGKILL");
   }
 }
-async function runCommand(command, root, spec, context, additions = {}) {
+function markedProcessPids(marker) {
+  const assignment = `${COMMAND_PROCESS_MARKER_ENV}=${marker}`;
+  if (process.platform === "linux") {
+    const matches2 = [];
+    let entries;
+    try {
+      entries = fs81.readdirSync("/proc");
+    } catch {
+      return matches2;
+    }
+    for (const entry of entries) {
+      if (!/^\d+$/.test(entry))
+        continue;
+      const pid = Number(entry);
+      if (pid === process.pid)
+        continue;
+      try {
+        const environment = fs81.readFileSync(path88.join("/proc", entry, "environ"), "utf8");
+        if (environment.split("\x00").includes(assignment))
+          matches2.push(pid);
+      } catch {}
+    }
+    return matches2;
+  }
+  if (process.platform === "darwin") {
+    try {
+      const output = execFileSync2("ps", ["eww", "-axo", "pid=,command="], {
+        encoding: "utf8",
+        maxBuffer: 16 * 1024 * 1024,
+        stdio: ["ignore", "pipe", "ignore"]
+      });
+      const matches2 = [];
+      for (const line of output.split(`
+`)) {
+        const match = line.match(/^\s*(\d+)\s+(.*)$/);
+        if (!match || !match[2].includes(assignment))
+          continue;
+        const pid = Number(match[1]);
+        if (Number.isInteger(pid) && pid !== process.pid)
+          matches2.push(pid);
+      }
+      return matches2;
+    } catch {
+      return [];
+    }
+  }
+  return [];
+}
+function processExists(pid) {
+  if (process.platform === "linux") {
+    try {
+      const stat = fs81.readFileSync(path88.join("/proc", String(pid), "stat"), "utf8");
+      const commandEnd = stat.lastIndexOf(")");
+      const state = commandEnd >= 0 ? stat.slice(commandEnd + 2, commandEnd + 3) : "";
+      if (state === "Z" || state === "X")
+        return false;
+    } catch (error2) {
+      if (error2.code === "ENOENT")
+        return false;
+    }
+  }
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (error2) {
+    return error2.code !== "ESRCH";
+  }
+}
+async function terminateCommandProcesses(child, marker, observedDescendants) {
+  if (child.pid !== undefined) {
+    for (const pid of descendantPids(child.pid))
+      observedDescendants.add(pid);
+  }
+  terminate(child);
+  const deadline = Date.now() + COMMAND_PROCESS_CLEANUP_MS;
+  while (true) {
+    for (const pid of markedProcessPids(marker))
+      observedDescendants.add(pid);
+    const active = [...observedDescendants].filter(processExists);
+    for (const pid of active) {
+      try {
+        process.kill(pid, "SIGKILL");
+      } catch {}
+    }
+    if (active.length === 0)
+      return;
+    if (Date.now() >= deadline) {
+      throw new BenchmarkPhaseError("command_cleanup_failed", "command descendants remained after bounded cleanup");
+    }
+    await new Promise((resolve) => setTimeout(resolve, COMMAND_PROCESS_POLL_MS));
+  }
+}
+async function runCommand(command, root, spec, context, options = {}) {
   const cwdRel = normalizedRootRelative(command.cwd);
   assertNoSymlinkTraversal(root, cwdRel);
   const cwd2 = path88.resolve(root, cwdRel);
@@ -79971,29 +80118,59 @@ async function runCommand(command, root, spec, context, additions = {}) {
   const remainingMs = context.deadline - Date.now();
   if (remainingMs <= 0)
     throw new BenchmarkPhaseError("phase_timeout", "phase budget expired");
-  const timeoutMs2 = Math.min(command.timeout_ms ?? remainingMs, remainingMs);
+  const reachesPhaseDeadline = command.timeout_ms === undefined || command.timeout_ms >= remainingMs;
+  const timeoutMs2 = reachesPhaseDeadline ? remainingMs : command.timeout_ms;
   const started = Date.now();
+  const commandMarker = crypto6.randomBytes(32).toString("hex");
   return await new Promise((resolve, reject2) => {
     const child = spawn4(command.argv[0], command.argv.slice(1), {
       cwd: cwd2,
-      env: buildEnvironment(spec, command.secret_env, additions),
+      env: buildEnvironment(spec, command.secret_env, {
+        ...options.additions,
+        [COMMAND_PROCESS_MARKER_ENV]: commandMarker
+      }),
       stdio: ["ignore", "pipe", "pipe"],
       detached: process.platform !== "win32"
     });
     const stdout = [];
     const stderr = [];
+    const observedDescendants = new Set;
     let captured = 0;
     let settled = false;
     let timer;
+    let observer;
+    let cleanup;
+    const observeDescendants = () => {
+      if (child.pid === undefined)
+        return;
+      for (const pid of descendantPids(child.pid))
+        observedDescendants.add(pid);
+    };
+    const cleanupProcesses = () => {
+      if (observer)
+        clearInterval(observer);
+      cleanup ??= terminateCommandProcesses(child, commandMarker, observedDescendants);
+      return cleanup;
+    };
+    const sanitizedOutput = (chunks) => redactCommandOutput(Buffer.concat(chunks), spec, [commandMarker]);
+    const sanitizedError = (error2) => {
+      if (!(error2 instanceof BenchmarkCommandExecutionError))
+        return error2;
+      return new BenchmarkCommandExecutionError(error2.code, error2.message, sanitizedOutput(stdout), sanitizedOutput(stderr), error2.durationMs);
+    };
     const fail = (error2) => {
       if (settled)
         return;
       settled = true;
       if (timer)
         clearTimeout(timer);
-      terminate(child);
-      reject2(error2);
+      cleanupProcesses().then(() => reject2(sanitizedError(error2)), reject2);
     };
+    if (process.platform !== "linux") {
+      observeDescendants();
+      observer = setInterval(observeDescendants, COMMAND_PROCESS_POLL_MS);
+      observer.unref();
+    }
     const capture = (target, chunk2) => {
       if (settled)
         return;
@@ -80005,29 +80182,48 @@ async function runCommand(command, root, spec, context, additions = {}) {
       }
       target.push(chunk2);
     };
+    const finish = (code, signal) => {
+      if (observer)
+        clearInterval(observer);
+      if (code === null) {
+        reject2(new BenchmarkCommandExecutionError("command_terminated", `command was terminated by ${signal ?? "an unknown signal"}: ${command.id}`, sanitizedOutput(stdout), sanitizedOutput(stderr), Date.now() - started));
+        return;
+      }
+      resolve({
+        exitCode: code,
+        stdout: sanitizedOutput(stdout),
+        stderr: sanitizedOutput(stderr),
+        durationMs: Date.now() - started,
+        redactions: [commandMarker]
+      });
+    };
     child.stdout?.on("data", (chunk2) => capture(stdout, chunk2));
     child.stderr?.on("data", (chunk2) => capture(stderr, chunk2));
     child.on("error", (error2) => {
-      fail(new BenchmarkPhaseError("command_start_failed", `failed to start ${command.id}: ${error2.message}`));
+      fail(new BenchmarkCommandExecutionError("command_start_failed", `failed to start ${command.id}: ${error2.message}`, Buffer.concat(stdout), Buffer.concat(stderr), Date.now() - started));
     });
     timer = setTimeout(() => {
-      fail(new BenchmarkPhaseError("command_timeout", `command timed out: ${command.id}`));
+      if (reachesPhaseDeadline) {
+        fail(new BenchmarkPhaseError("phase_timeout", "phase budget expired"));
+        return;
+      }
+      fail(new BenchmarkCommandExecutionError("command_timeout", `command timed out: ${command.id}`, Buffer.concat(stdout), Buffer.concat(stderr), Date.now() - started));
     }, timeoutMs2);
-    child.on("close", (code, signal) => {
+    child.on("exit", (code, signal) => {
       if (settled)
         return;
       settled = true;
       clearTimeout(timer);
-      if (code === null) {
-        reject2(new BenchmarkPhaseError("command_terminated", `command was terminated by ${signal ?? "an unknown signal"}: ${command.id}`));
-        return;
+      if (options.descendantCleanup === "always" || code === null || code !== 0) {
+        cleanupProcesses().then(() => finish(code, signal), reject2);
+      } else {
+        if (observer)
+          clearInterval(observer);
+        child.unref();
+        child.stdout?.unref?.();
+        child.stderr?.unref?.();
+        finish(code, signal);
       }
-      resolve({
-        exitCode: code,
-        stdout: Buffer.concat(stdout),
-        stderr: Buffer.concat(stderr),
-        durationMs: Date.now() - started
-      });
     });
   });
 }
@@ -80100,11 +80296,14 @@ async function executeHydrate(spec, context) {
       throw error2;
     }
   }
+  removePrivateHydrationSourcesBeforeSetup(spec, context);
   for (const command of spec.setup_commands) {
     assertBudget(context);
     let result2;
     try {
-      result2 = await runCommand(command, spec.workspace_root, spec, context);
+      result2 = await runCommand(command, spec.workspace_root, spec, context, {
+        descendantCleanup: "failure_only"
+      });
     } catch (error2) {
       context.steps.push({
         id: command.id,
@@ -80155,7 +80354,6 @@ async function executeHydrate(spec, context) {
   }
   outputs.splice(0, outputs.length, ...finalOutputs);
   context.outputs.splice(0, context.outputs.length, ...finalOutputs);
-  removeRemoteHydrationInputs(spec);
   assertBudget(context);
   return outputs;
 }
@@ -80220,6 +80418,49 @@ async function workspaceManifest(spec, context) {
   }
   return records.sort((a3, b4) => a3.path.localeCompare(b4.path));
 }
+async function verifyWorkspaceManifestUnchanged(expected, spec, context) {
+  await verifyRecordsUnchanged(expected, spec);
+  const actual = await workspaceManifest(spec, context);
+  if (JSON.stringify(actual) !== JSON.stringify(expected)) {
+    throw new BenchmarkPhaseError("evidence_tampered", "candidate workspace changed during read-only evaluation");
+  }
+}
+function treeBytes(root, context) {
+  let totalBytes = 0;
+  const stack = [path88.resolve(root)];
+  while (stack.length > 0) {
+    const directory = stack.pop();
+    const entries = fs81.readdirSync(directory, { withFileTypes: true });
+    for (const entry of entries) {
+      assertBudget(context);
+      const candidate = path88.join(directory, entry.name);
+      const stat = fs81.lstatSync(candidate);
+      if (stat.isDirectory()) {
+        stack.push(candidate);
+        continue;
+      }
+      if (stat.isFile()) {
+        totalBytes += stat.size;
+        continue;
+      }
+      if (stat.isSymbolicLink()) {
+        totalBytes += Buffer.byteLength(fs81.readlinkSync(candidate));
+        continue;
+      }
+      throw new BenchmarkPhaseError("unsafe_path", `evaluate output contains an unsupported filesystem entry: ${path88.relative(root, candidate)}`);
+    }
+  }
+  return totalBytes;
+}
+function assertEvaluateOutputBudget(spec, context, additionalBytes = 0) {
+  const actual = treeBytes(spec.logs_root, context) + additionalBytes;
+  if (actual > spec.workspace_limits.max_total_bytes) {
+    throw new BenchmarkPhaseError("workspace_limit_exceeded", "evaluate output bytes exceed the workspace byte limit", {
+      actual,
+      max_total_bytes: spec.workspace_limits.max_total_bytes
+    });
+  }
+}
 function candidateGlob(pattern) {
   let source = "^";
   for (let index = 0;index < pattern.length; index += 1) {
@@ -80250,6 +80491,8 @@ function manifestDirectories(manifest, context) {
   const directories = new Set;
   for (const entry of manifest) {
     assertBudget(context);
+    if (entry.kind === "symlink")
+      continue;
     let current = path88.posix.dirname(entry.path);
     while (current !== ".") {
       assertBudget(context);
@@ -80267,9 +80510,8 @@ function candidateOutputFiles(output, manifest, context) {
       assertBudget(context);
       if (!matcher.test(entry.path))
         continue;
-      if (entry.kind === "symlink") {
-        throw new BenchmarkPhaseError("unsafe_path", `candidate output ${output.id} matches a symlink`);
-      }
+      if (entry.kind === "symlink")
+        continue;
       matched.push(entry);
     }
     return {
@@ -80277,12 +80519,6 @@ function candidateOutputFiles(output, manifest, context) {
       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);
@@ -80297,9 +80533,8 @@ function candidateOutputFiles(output, manifest, context) {
       assertBudget(context);
       if (!entry.path.startsWith(prefix))
         continue;
-      if (entry.kind === "symlink") {
-        throw new BenchmarkPhaseError("unsafe_path", `candidate output ${output.id} contains a symlink`);
-      }
+      if (entry.kind === "symlink")
+        continue;
       selected.set(entry.path, entry);
     }
   }
@@ -80332,7 +80567,7 @@ async function copyCandidateOutput(output, manifest, spec, context) {
   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-artifacts", output.id, 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}`);
     }
@@ -80345,21 +80580,29 @@ async function copyCandidateOutput(output, manifest, spec, context) {
   }
   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}`);
-    }
-  }
+async function copyFrozenWorkspace(manifest, spec, context, sourceRoot = spec.workspace_root, symlinkPolicy = "preserve") {
   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 source = path88.resolve(sourceRoot, safeRelPath(frozenFile.path));
     const destination = path88.resolve(destinationRoot, safeRelPath(frozenFile.path));
-    await atomicCopy(source, destination, frozenFile.mode, spec.workspace_root);
+    if (frozenFile.kind === "symlink") {
+      const stat = fs81.lstatSync(source);
+      const target = stat.isSymbolicLink() ? fs81.readlinkSync(source) : null;
+      if (target === null || Buffer.byteLength(target) !== frozenFile.size || sha256(target) !== frozenFile.sha256) {
+        throw new BenchmarkPhaseError("evidence_tampered", `workspace changed after it was frozen: ${frozenFile.path}`);
+      }
+      const resolvedTarget = path88.resolve(path88.dirname(source), target);
+      if (symlinkPolicy === "contained_relative_only" && (path88.isAbsolute(target) || !isWithin(sourceRoot, resolvedTarget))) {
+        continue;
+      }
+      fs81.mkdirSync(path88.dirname(destination), { recursive: true, mode: 448 });
+      fs81.symlinkSync(target, destination);
+      continue;
+    }
+    await atomicCopy(source, destination, frozenFile.mode, sourceRoot);
     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}`);
@@ -80367,13 +80610,52 @@ async function copyFrozenWorkspace(manifest, spec, context) {
   }
   return destinationRoot;
 }
-function evaluatorTestsPath(evaluator, spec) {
+async function copyEvaluatorTests(spec, evaluator, context) {
+  const destinationRoot = fs81.mkdtempSync(path88.join(os16.tmpdir(), "brainbase-benchmark-tests-"));
+  fs81.chmodSync(destinationRoot, 448);
+  context.temporaryRoots.add(destinationRoot);
+  const sourceRoot = evaluatorTestsPath(evaluator, spec.tests_root);
+  const relativeRoot = evaluator.tests_path ? normalizedRootRelative(evaluator.tests_path) : ".";
+  const destinationStart = relativeRoot === "." ? destinationRoot : path88.resolve(destinationRoot, relativeRoot);
+  fs81.mkdirSync(destinationStart, { recursive: true, mode: 448 });
+  const stack = [{ source: sourceRoot, destination: destinationStart }];
+  while (stack.length > 0) {
+    assertBudget(context);
+    const current = stack.pop();
+    const entries = fs81.readdirSync(current.source, { withFileTypes: true }).sort((left, right) => left.name.localeCompare(right.name));
+    for (const entry of entries) {
+      assertBudget(context);
+      const source = path88.join(current.source, entry.name);
+      const destination = path88.join(current.destination, entry.name);
+      const stat = fs81.lstatSync(source);
+      if (stat.isSymbolicLink()) {
+        throw new BenchmarkPhaseError("unsafe_path", `evaluator reference tree contains a symlink: ${path88.relative(spec.tests_root, source)}`);
+      }
+      if (stat.isDirectory()) {
+        fs81.mkdirSync(destination, { recursive: true, mode: stat.mode & 511 });
+        stack.push({ source, destination });
+        continue;
+      }
+      if (!stat.isFile()) {
+        throw new BenchmarkPhaseError("unsafe_path", `evaluator reference tree contains an unsupported filesystem entry: ${path88.relative(spec.tests_root, source)}`);
+      }
+      await atomicCopy(source, destination, stat.mode & 511, spec.tests_root);
+      const sourceRecord = await recordFile(spec.tests_root, source, "tests");
+      const copiedRecord = await recordFile(destinationRoot, destination, "tests");
+      if (copiedRecord.path !== sourceRecord.path || copiedRecord.sha256 !== sourceRecord.sha256 || copiedRecord.size !== sourceRecord.size || copiedRecord.mode !== sourceRecord.mode) {
+        throw new BenchmarkPhaseError("evidence_tampered", `evaluator reference changed while it was copied: ${sourceRecord.path}`);
+      }
+    }
+  }
+  return destinationRoot;
+}
+function evaluatorTestsPath(evaluator, testsRoot) {
   if (!evaluator.tests_path)
-    return spec.tests_root;
+    return testsRoot;
   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)) {
+  assertNoSymlinkTraversal(testsRoot, relative);
+  const candidate = path88.resolve(testsRoot, relative);
+  if (!isWithin(testsRoot, candidate) || !fs81.existsSync(candidate)) {
     throw new BenchmarkPhaseError("invalid_tests_path", `evaluator tests_path does not exist: ${evaluator.tests_path}`);
   }
   const stat = fs81.lstatSync(candidate);
@@ -80398,8 +80680,39 @@ var CriterionResultSchema = exports_external.object({
 });
 var CriteriaResultFileSchema = exports_external.object({
   criteria: exports_external.array(CriterionResultSchema).max(MAX_CRITERIA_PER_EVALUATOR)
+}).strict().superRefine((value, context) => {
+  if (Buffer.byteLength(JSON.stringify(value)) > MAX_CRITERIA_RESULT_PAYLOAD_BYTES) {
+    context.addIssue({
+      code: exports_external.ZodIssueCode.custom,
+      path: ["criteria"],
+      message: "criterion result exceeds the 20 MiB aggregate payload limit"
+    });
+  }
+});
+var JudgeErrorSchema = exports_external.object({
+  code: exports_external.enum([
+    "judge_provider_quota_exhausted",
+    "judge_rate_limited",
+    "judge_provider_authentication_failed",
+    "judge_provider_unavailable",
+    "judge_request_failed"
+  ]),
+  message: exports_external.string().min(1).max(500)
 }).strict();
-function readCriterionResults(resultPath, criterionKeys, allowedEvidenceIds) {
+function structuredJudgeError(stderr) {
+  for (const line of stderr.toString("utf8").split(/\r?\n/).reverse()) {
+    const marker = line.indexOf(JUDGE_ERROR_PREFIX);
+    if (marker < 0)
+      continue;
+    try {
+      const parsed = JudgeErrorSchema.safeParse(JSON.parse(line.slice(marker + JUDGE_ERROR_PREFIX.length)));
+      if (parsed.success)
+        return parsed.data;
+    } catch {}
+  }
+  return null;
+}
+function readCriterionResults(resultPath, criterionKeys, allowedEvidenceIds, spec, additionalSecrets = []) {
   let opened;
   try {
     opened = openRegularFileNoFollow(resultPath, "criterion result", path88.dirname(resultPath));
@@ -80411,7 +80724,7 @@ function readCriterionResults(resultPath, criterionKeys, allowedEvidenceIds) {
   }
   try {
     if (opened.stat.size > MAX_CRITERIA_RESULT_BYTES) {
-      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result exceeds the 1 MiB limit");
+      throw new BenchmarkPhaseError("invalid_criterion_result", "criterion result exceeds the 24 MiB file limit");
     }
     let raw;
     try {
@@ -80443,7 +80756,13 @@ function readCriterionResults(resultPath, criterionKeys, allowedEvidenceIds) {
     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));
+    return criterionKeys.map((key2) => {
+      const criterion = byKey.get(key2);
+      return {
+        ...criterion,
+        explanation: redactCommandOutput(Buffer.from(criterion.explanation), spec, additionalSecrets).toString("utf8")
+      };
+    });
   } finally {
     fs81.closeSync(opened.fd);
   }
@@ -80466,7 +80785,7 @@ function eventType(event) {
   }
   return;
 }
-async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, frozenEvidence, context) {
+async function evaluateOne(evaluator, spec, manifest, frozenWorkspaceRoot, finalOutput, trajectory, frozenEvidence, context) {
   const started = Date.now();
   const base2 = {
     id: evaluator.id,
@@ -80501,16 +80820,33 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
         cwd: ".",
         secret_env: []
       }, spec.tests_root, spec, context, {
-        BB_REGEX_INPUT: frozenEvidence.finalOutputPath,
-        BB_REGEX_PATTERN_B64: Buffer.from(assertion.pattern).toString("base64"),
-        BB_REGEX_FLAGS: assertion.flags
+        additions: {
+          BB_REGEX_INPUT: frozenEvidence.finalOutputPath,
+          BB_REGEX_PATTERN_B64: Buffer.from(assertion.pattern).toString("base64"),
+          BB_REGEX_FLAGS: assertion.flags
+        },
+        descendantCleanup: "always"
       });
       if (regexResult.exitCode === 2) {
         throw new BenchmarkPhaseError("invalid_evaluator", `invalid regex in evaluator: ${evaluator.id}`);
       }
       verdict = regexResult.exitCode === 0;
     }
-    return { ...base2, status: verdict ? "passed" : "failed", verdict, duration_ms: Date.now() - started };
+    return {
+      ...base2,
+      status: verdict ? "passed" : "failed",
+      verdict,
+      duration_ms: Date.now() - started,
+      details: {
+        operator: assertion.operator,
+        actual_size_bytes: Buffer.byteLength(finalOutput),
+        ...assertion.operator === "exact" ? {
+          expected_size_bytes: Buffer.byteLength(assertion.expected),
+          actual_sha256: sha256(finalOutput),
+          expected_sha256: sha256(assertion.expected)
+        } : {}
+      }
+    };
   }
   if (evaluator.type === "trajectory_assertion") {
     const count = trajectory.filter((event) => eventType(event) === evaluator.event_type).length;
@@ -80525,8 +80861,8 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
   }
   if (evaluator.type === "workspace_assertion") {
     const relative = workspaceRel(evaluator.path);
-    assertNoSymlinkTraversal(spec.workspace_root, relative);
-    const candidate = path88.resolve(spec.workspace_root, relative);
+    assertNoSymlinkTraversal(frozenWorkspaceRoot, relative);
+    const candidate = path88.resolve(frozenWorkspaceRoot, relative);
     let stat = null;
     try {
       stat = fs81.lstatSync(candidate);
@@ -80540,15 +80876,17 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
     }
     const exists2 = stat !== null;
     let verdict = false;
+    let actualSha256;
     if (evaluator.assertion.operator === "exists")
       verdict = exists2;
     if (evaluator.assertion.operator === "not_exists")
       verdict = !exists2;
     if (evaluator.assertion.operator === "sha256") {
       if (stat?.isFile()) {
-        const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, spec.workspace_root);
+        const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, frozenWorkspaceRoot);
         try {
-          verdict = await sha256OfDescriptor(opened.fd) === evaluator.assertion.expected;
+          actualSha256 = await sha256OfDescriptor(opened.fd);
+          verdict = actualSha256 === evaluator.assertion.expected;
         } finally {
           fs81.closeSync(opened.fd);
         }
@@ -80556,7 +80894,7 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
     }
     if (evaluator.assertion.operator === "contains") {
       if (stat?.isFile()) {
-        const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, spec.workspace_root);
+        const opened = openRegularFileNoFollow(candidate, `workspace assertion ${relative}`, frozenWorkspaceRoot);
         try {
           verdict = readDescriptor(opened.fd).toString("utf8").includes(evaluator.assertion.expected);
         } finally {
@@ -80564,15 +80902,36 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
         }
       }
     }
-    return { ...base2, status: verdict ? "passed" : "failed", verdict, duration_ms: Date.now() - started };
+    return {
+      ...base2,
+      status: verdict ? "passed" : "failed",
+      verdict,
+      duration_ms: Date.now() - started,
+      details: {
+        operator: evaluator.assertion.operator,
+        path: relative,
+        exists: exists2,
+        ...stat ? {
+          actual_kind: stat.isFile() ? "file" : stat.isDirectory() ? "directory" : "other",
+          actual_size_bytes: stat.size
+        } : {},
+        ...actualSha256 ? { actual_sha256: actualSha256 } : {}
+      }
+    };
   }
   let isolatedWorkspace;
+  let isolatedTestsRoot;
   let privateResultRoot;
+  const usesLiveWorkspace = evaluator.workspace_mode === "read_only";
   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 evaluatorWorkspace = usesLiveWorkspace ? spec.workspace_root : await copyFrozenWorkspace(manifest, spec, context, frozenWorkspaceRoot, "contained_relative_only");
+    if (!usesLiveWorkspace)
+      isolatedWorkspace = evaluatorWorkspace;
+    else
+      await verifyWorkspaceManifestUnchanged(manifest, spec, context);
+    isolatedTestsRoot = await copyEvaluatorTests(spec, evaluator, context);
+    const testsPath = evaluatorTestsPath(evaluator, isolatedTestsRoot);
+    const commandRoot = evaluator.root === "workspace" ? evaluatorWorkspace : isolatedTestsRoot;
     const command = { id: evaluator.id, ...evaluator.command };
     const environment = {
       BRAINBASE_BENCHMARK_WORKSPACE: evaluatorWorkspace,
@@ -80589,14 +80948,23 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
       criterionResultPath = path88.join(privateResultRoot, "result.json");
       environment.BRAINBASE_BENCHMARK_CRITERIA_RESULT = criterionResultPath;
     }
-    const result2 = await runCommand(command, commandRoot, spec, context, environment);
+    const result2 = await runCommand(command, commandRoot, spec, context, {
+      additions: environment,
+      descendantCleanup: "always"
+    });
     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;
+      criterionResults = evaluator.criterion_keys && criterionResultPath ? readCriterionResults(criterionResultPath, evaluator.criterion_keys, evaluator.allowed_evidence_ids ?? [], spec, result2.redactions) : undefined;
     } catch (error2) {
-      const normalized = stableError(error2);
+      let normalized = stableError(error2);
+      if (normalized?.code === "missing_criterion_result" && result2.exitCode !== 0) {
+        normalized = structuredJudgeError(result2.stderr) ?? {
+          code: "command_terminated",
+          message: "sandbox evaluator terminated before writing its criterion result"
+        };
+      }
       return {
         ...base2,
         status: "errored",
@@ -80623,11 +80991,17 @@ async function evaluateOne(evaluator, spec, manifest, finalOutput, trajectory, f
       ...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);
+    try {
+      if (usesLiveWorkspace) {
+        await verifyWorkspaceManifestUnchanged(manifest, spec, context);
+      }
+    } finally {
+      for (const temporary of [privateResultRoot, isolatedTestsRoot, isolatedWorkspace]) {
+        if (!temporary)
+          continue;
+        fs81.rmSync(temporary, { recursive: true, force: true });
+        context.temporaryRoots.delete(temporary);
+      }
     }
   }
 }
@@ -80672,6 +81046,7 @@ async function executeEvaluate(spec, context) {
   ];
   outputs.push(...frozenEvidenceRecords);
   context.outputs.push(...frozenEvidenceRecords);
+  assertEvaluateOutputBudget(spec, context);
   assertBudget(context);
   const manifest = await workspaceManifest(spec, context);
   const manifestPath2 = path88.join(spec.logs_root, "candidate-workspace-manifest.json");
@@ -80684,6 +81059,7 @@ async function executeEvaluate(spec, context) {
   const manifestRecord = await recordFile(spec.logs_root, manifestPath2, "logs");
   outputs.push(manifestRecord);
   context.outputs.push(manifestRecord);
+  assertEvaluateOutputBudget(spec, context);
   for (const artifactRelInput of spec.candidate_artifacts) {
     const artifactRel = workspaceRel(artifactRelInput);
     assertNoSymlinkTraversal(spec.workspace_root, artifactRel);
@@ -80701,11 +81077,13 @@ async function executeEvaluate(spec, context) {
     outputs.push(artifact);
     context.outputs.push(artifact);
   }
+  assertEvaluateOutputBudget(spec, context);
   for (const candidateOutput of spec.candidate_outputs) {
     const copied = await copyCandidateOutput(candidateOutput, manifest, spec, context);
     outputs.push(...copied);
     context.outputs.push(...copied);
   }
+  assertEvaluateOutputBudget(spec, context);
   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");
@@ -80719,23 +81097,23 @@ async function executeEvaluate(spec, context) {
     const archiveRecord = await recordFile(spec.logs_root, archive, "logs");
     outputs.push(archiveRecord);
     context.outputs.push(archiveRecord);
+    assertEvaluateOutputBudget(spec, context);
   }
   await verifyRecordsUnchanged(manifest, spec);
+  const frozenWorkspaceRoot = await copyFrozenWorkspace(manifest, spec, context, spec.workspace_root);
   for (const reference of spec.references) {
     const referenceOutputs = await copyMaterial(reference, spec.staging_root, spec.tests_root, "tests", false, context);
     outputs.push(...referenceOutputs);
     context.outputs.push(...referenceOutputs);
   }
-  const frozenOutputCount = context.outputs.length;
   const evaluators = [];
   const evaluatorOrder = new Map(spec.evaluators.map((evaluator, index) => [evaluator.id, index]));
   const executionOrder = [...spec.evaluators].sort((left, right) => Number(left.type === "sandbox_command") - Number(right.type === "sandbox_command"));
   for (const evaluator of executionOrder) {
     assertBudget(context);
     const started = Date.now();
-    let requiredResultError;
     try {
-      const evaluated = await evaluateOne(evaluator, spec, manifest, finalOutput.buffer.toString("utf8"), trajectory, { finalOutputPath, trajectoryPath }, context);
+      const evaluated = await evaluateOne(evaluator, spec, manifest, frozenWorkspaceRoot, finalOutput.buffer.toString("utf8"), trajectory, { finalOutputPath, trajectoryPath }, context);
       assertBudget(context);
       evaluators.push(evaluated);
       context.evaluators.push(evaluated);
@@ -80747,13 +81125,14 @@ 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);
+      let stdout;
+      let stderr;
+      if (error2 instanceof BenchmarkCommandExecutionError) {
+        stdout = await writeLog(spec.logs_root, `${evaluator.id}.stdout.log`, error2.stdout, spec);
+        stderr = await writeLog(spec.logs_root, `${evaluator.id}.stderr.log`, error2.stderr, spec);
+      }
       const errored = {
         id: evaluator.id,
         type: evaluator.type,
@@ -80763,26 +81142,38 @@ async function executeEvaluate(spec, context) {
         verdict: null,
         engine: "brainbase-cli",
         engine_version: VERSION,
-        duration_ms: Date.now() - started,
+        duration_ms: error2 instanceof BenchmarkCommandExecutionError ? error2.durationMs : Date.now() - started,
         details: {
           error_code: normalized?.code ?? "phase_failed",
           error_message: normalized?.message ?? "evaluator execution failed"
-        }
+        },
+        ...stdout ? { stdout } : {},
+        ...stderr ? { stderr } : {}
       };
       evaluators.push(errored);
       context.evaluators.push(errored);
-      if (evaluator.required)
+      if (stdout) {
+        outputs.push(stdout);
+        context.outputs.push(stdout);
+      }
+      if (stderr) {
+        outputs.push(stderr);
+        context.outputs.push(stderr);
+      }
+      if (!(error2 instanceof BenchmarkCommandExecutionError))
         throw error2;
       assertBudget(context);
     }
-    if (requiredResultError)
-      throw requiredResultError;
+    if (evaluator.type === "sandbox_command") {
+      assertEvaluateOutputBudget(spec, context);
+    }
   }
-  await verifyRecordsUnchanged([...context.inputs, ...context.outputs.slice(0, frozenOutputCount)], spec);
+  await verifyRecordsUnchanged([...context.inputs, ...context.outputs], spec);
   if (!verifyOwnedDirectory(spec.tests_root, "tests", spec) || !verifyOwnedDirectory(spec.logs_root, "logs", spec)) {
     throw new BenchmarkPhaseError("evidence_tampered", "benchmark evaluator changed an owned phase directory marker");
   }
   assertBudget(context);
+  assertEvaluateOutputBudget(spec, context);
   evaluators.sort((left, right) => evaluatorOrder.get(left.id) - evaluatorOrder.get(right.id));
   return { outputs, evaluators };
 }
@@ -80809,6 +81200,67 @@ function stableError(error2) {
     message: error2 instanceof Error ? error2.message : "unknown benchmark phase failure"
   };
 }
+function compactOversizedResult(result2) {
+  const actualBytes = Buffer.byteLength(`${JSON.stringify(result2, null, 2)}
+`);
+  if (actualBytes <= MAX_PHASE_RESULT_BYTES)
+    return result2;
+  return {
+    schema_version: result2.schema_version,
+    cli_version: result2.cli_version,
+    phase: result2.phase,
+    attempt_id: result2.attempt_id,
+    phase_id: result2.phase_id,
+    spec_digest: result2.spec_digest,
+    status: "failed",
+    started_at: result2.started_at,
+    completed_at: result2.completed_at,
+    duration_ms: result2.duration_ms,
+    steps: [],
+    inputs: [],
+    outputs: [],
+    error: {
+      code: "result_too_large",
+      message: "benchmark phase result exceeds the 32 MiB contract limit",
+      details: {
+        actual_bytes: actualBytes,
+        max_bytes: MAX_PHASE_RESULT_BYTES,
+        step_count: result2.steps.length,
+        input_count: result2.inputs.length,
+        output_count: result2.outputs.length,
+        evaluator_count: result2.evaluators?.length ?? 0
+      }
+    }
+  };
+}
+function readCachedPhaseResult(resultPath) {
+  let opened;
+  try {
+    opened = openRegularFileNoFollow(resultPath, "cached benchmark phase result");
+  } catch {
+    return;
+  }
+  try {
+    if (opened.stat.size > MAX_PHASE_RESULT_BYTES)
+      return;
+    const chunks = [];
+    let offset = 0;
+    while (offset <= MAX_PHASE_RESULT_BYTES) {
+      const buffer = Buffer.alloc(Math.min(64 * 1024, MAX_PHASE_RESULT_BYTES + 1 - offset));
+      const bytesRead = fs81.readSync(opened.fd, buffer, 0, buffer.length, offset);
+      if (bytesRead === 0) {
+        return JSON.parse(Buffer.concat(chunks, offset).toString("utf8"));
+      }
+      chunks.push(buffer.subarray(0, bytesRead));
+      offset += bytesRead;
+    }
+    return;
+  } catch {
+    return;
+  } finally {
+    fs81.closeSync(opened.fd);
+  }
+}
 function rawIdentity(value) {
   if (!value || typeof value !== "object") {
     return { phase: "unknown", attemptId: null, phaseId: null };
@@ -80881,13 +81333,9 @@ function prepareBenchmarkPhaseInvocation(specPathInput, resultPathInput, expecte
       timeout_ms: spec.budget.timeout_ms
     };
     let cachedResult;
-    if (fs81.existsSync(resultPath)) {
-      try {
-        const cached2 = JSON.parse(fs81.readFileSync(resultPath, "utf8"));
-        if (cached2.schema_version === SCHEMA_VERSION && cached2.phase === spec.phase && cached2.attempt_id === spec.attempt_id && cached2.phase_id === spec.phase_id && cached2.spec_digest === digest && cached2.status === "succeeded") {
-          cachedResult = cached2;
-        }
-      } catch {}
+    const cached2 = readCachedPhaseResult(resultPath);
+    if (cached2?.schema_version === SCHEMA_VERSION && cached2.phase === spec.phase && cached2.attempt_id === spec.attempt_id && cached2.phase_id === spec.phase_id && cached2.spec_digest === digest && cached2.status === "succeeded") {
+      cachedResult = cached2;
     }
     if (cachedResult) {
       if (spec.phase === "hydrate") {
@@ -80904,7 +81352,7 @@ function prepareBenchmarkPhaseInvocation(specPathInput, resultPathInput, expecte
   } catch (error2) {
     return {
       ok: false,
-      result: {
+      result: compactOversizedResult({
         schema_version: SCHEMA_VERSION,
         cli_version: VERSION,
         phase: identity2.phase === "unknown" ? expectedPhase : identity2.phase,
@@ -80919,7 +81367,7 @@ function prepareBenchmarkPhaseInvocation(specPathInput, resultPathInput, expecte
         inputs: [],
         outputs: [],
         error: stableError(error2)
-      }
+      })
     };
   }
 }
@@ -80968,6 +81416,7 @@ async function runBenchmarkPhase(specPathInput, resultPathInput, expectedPhase,
   let status = "failed";
   let error2;
   let context;
+  let validatedSpec;
   let resultPathValidated = false;
   try {
     let bytes;
@@ -80987,6 +81436,7 @@ async function runBenchmarkPhase(specPathInput, resultPathInput, expectedPhase,
     }
     identity2 = rawIdentity(raw);
     const spec = BenchmarkSpecSchema.parse(raw);
+    validatedSpec = spec;
     if (expectedPhase && spec.phase !== expectedPhase) {
       throw new BenchmarkPhaseError("phase_mismatch", `the ${expectedPhase} command cannot execute a ${spec.phase} spec`);
     }
@@ -80997,13 +81447,9 @@ async function runBenchmarkPhase(specPathInput, resultPathInput, expectedPhase,
     }
     resultPathValidated = true;
     let cachedResult;
-    if (fs81.existsSync(resultPath)) {
-      try {
-        const cached2 = JSON.parse(fs81.readFileSync(resultPath, "utf8"));
-        if (cached2.schema_version === SCHEMA_VERSION && cached2.phase === spec.phase && cached2.attempt_id === spec.attempt_id && cached2.phase_id === spec.phase_id && cached2.spec_digest === digest && cached2.status === "succeeded") {
-          cachedResult = cached2;
-        }
-      } catch {}
+    const cached2 = readCachedPhaseResult(resultPath);
+    if (cached2?.schema_version === SCHEMA_VERSION && cached2.phase === spec.phase && cached2.attempt_id === spec.attempt_id && cached2.phase_id === spec.phase_id && cached2.spec_digest === digest && cached2.status === "succeeded") {
+      cachedResult = cached2;
     }
     if (cachedResult) {
       if (spec.phase === "hydrate") {
@@ -81046,7 +81492,7 @@ async function runBenchmarkPhase(specPathInput, resultPathInput, expectedPhase,
       fs81.rmSync(temporary, { recursive: true, force: true });
     }
   }
-  const result2 = {
+  let result2 = {
     schema_version: SCHEMA_VERSION,
     cli_version: VERSION,
     phase: identity2.phase === "unknown" && expectedPhase ? expectedPhase : identity2.phase,
@@ -81064,8 +81510,22 @@ async function runBenchmarkPhase(specPathInput, resultPathInput, expectedPhase,
     ...error2 ? { error: error2 } : {}
   };
   if (!resultPathValidated || !context?.logsOwned) {
+    result2 = compactOversizedResult(result2);
     return { exitCode: status === "succeeded" ? 0 : 1, result: result2 };
   }
+  if (status === "succeeded" && validatedSpec?.phase === "evaluate") {
+    try {
+      const resultBytes = Buffer.byteLength(`${JSON.stringify(result2, null, 2)}
+`);
+      assertEvaluateOutputBudget(validatedSpec, context, resultBytes);
+    } catch (budgetError) {
+      status = "failed";
+      result2.status = "failed";
+      result2.error = stableError(budgetError);
+    }
+  }
+  result2 = compactOversizedResult(result2);
+  status = result2.status;
   try {
     writeJsonAtomic(resultPath, result2);
   } catch (writeError) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@brainbase-labs/cli",
-  "version": "0.25.0-eng1209.1",
+  "version": "0.25.0-eng1209.11",
   "description": "Pack, share, and install agent templates across harnesses (Claude Code, Codex, ...).",
   "type": "module",
   "bin": {