@workbench-ai/workbench-core 0.0.48 → 0.0.50

package/dist/index.js

@@ -4,19 +4,19 @@ import path from "node:path";
 import { fileURLToPath } from "node:url";
 import YAML from "yaml";
 import { adapterCommandName, assertWorkbenchAdapterOperationResultOk, collectWorkbenchAdapterAuthRequirements, parseWorkbenchAdapterManifest, readWorkbenchAdapterOperationResult, WORKBENCH_RUNTIME_CONTROL_TOKEN_ENV, WORKBENCH_RUNTIME_CONTROL_URL_ENV, workbenchAdapterOperationCommand, workbenchAdapterOperationExecutor, workbenchAdapterOperationResultPath, } from "@workbench-ai/workbench-protocol";
-import { BENCHMARK_SPEC_FILE, engineCasePrivateFiles, engineCaseFilesForRuntimeInput, engineCasePublicFiles, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml as resolveWorkbenchResolvedSourceYamlInternal, validateWorkbenchResolvedSourceYaml as validateWorkbenchResolvedSourceYamlInternal, isWorkbenchSubjectManifestPath, } from "./generic-spec.js";
+import { BENCHMARK_SPEC_FILE, engineCasePrivateFiles, engineCaseFilesForRuntimeInput, engineCasePublicFiles, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml as resolveWorkbenchResolvedSourceYamlInternal, validateWorkbenchResolvedSourceYaml as validateWorkbenchResolvedSourceYamlInternal, isWorkbenchCandidateManifestPath, } from "./generic-spec.js";
 import { attachSandboxMetadataToJob, createWorkbenchSandboxFileStore, isSurfaceSnapshotFile, readWorkbenchExecutionSpec, } from "./sandbox-inputs.js";
 import { asRuntimeRecord, importNodeModule, isJsonPayload, jsonRecord, nodeBuiltin, quoteShellArg, resolveWorkbenchWorkerId, } from "./runtime-utils.js";
 import { createWorkbenchExecutionCapability, createWorkbenchSandboxAllocation, collectExecutionCapabilityScopeIssues, collectSandboxAllocationScopeIssues, collectSandboxHandleScopeIssues, assertSandboxBackendSupportsNetworkPolicy, executeValidatedSandboxExecution, } from "./sandbox-plane.js";
 import { createSandboxBackendPlaneForProvider, } from "./sandbox-backends/index.js";
-import { applyWorkbenchSubjectPatch } from "./subject-patch.js";
+import { applyWorkbenchCandidatePatch } from "./candidate-patch.js";
 import { assignUsageRole, completeUsageSummary, mergeUsageSummaries, normalizeUsageSummary, usageStats, } from "./execution-usage.js";
 import { traceFilePaths, workbenchTraceExecutionDirectory, } from "./trace-files.js";
 import { engineCaseForCase, } from "./execution-jobs.js";
 import { createWorkbenchExecutionEventPublisher, publishCommandStepEvent, } from "./execution-events.js";
 import { readWorkbenchExecutionPurpose } from "./execution-evidence.js";
 import { adapterAuthEnv, localWorkbenchAdapterAuthStore, normalizeWorkbenchAdapterAuthTarget, sanitizeWorkbenchAdapterAuthBundle, } from "./adapter-auth.js";
-export { BENCHMARK_SPEC_FILE, DEFAULT_EXECUTION_RESOURCES, engineCasePrivateFiles, engineCaseFilesForRuntimeInput, engineCasePublicFiles, engineResolveInvocationForSpec, engineResolveBindingForSpec, engineResolveBindingForSourceYaml, isWorkbenchSubjectManifestPath, parseWorkbenchSourceFiles, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml, resolveWorkbenchSourceFiles, runtimeNetwork, runtimeResources, serializeWorkbenchResolvedSourceYaml, validateWorkbenchResolvedSourceYaml, } from "./generic-spec.js";
+export { BENCHMARK_SPEC_FILE, CANDIDATE_SPEC_FILE, DEFAULT_EXECUTION_RESOURCES, engineCasePrivateFiles, engineCaseFilesForRuntimeInput, engineCasePublicFiles, engineResolveInvocationForSpec, engineResolveBindingForSpec, engineResolveBindingForSourceYaml, isWorkbenchCandidateManifestPath, parseWorkbenchSourceFiles, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml, resolveWorkbenchSourceFiles, runtimeNetwork, runtimeResources, serializeWorkbenchResolvedSourceYaml, validateWorkbenchResolvedSourceYaml, } from "./generic-spec.js";
 export { composeRuntimeDockerfileWithAdapterInstallers, } from "./runtime-dockerfile.js";
 export { adapterCommandName, cloneWorkbenchAdapterManifest, collectWorkbenchAdapterAuthRequirements, collectWorkbenchAdapterInvocations, parseWorkbenchAdapterManifest, workbenchAdapterManifestRequiresAuth, workbenchAdapterManifestSupportsOperation, workbenchAdapterOperationCommand, workbenchAdapterOperationExecutor, withDefaultWorkbenchAdapterAuth, withDefaultWorkbenchAdapterAuthProfiles, } from "@workbench-ai/workbench-protocol";
 export { adapterAuthEnv, createWorkbenchAdapterAuthBundle, defaultWorkbenchAdapterAuthStoreRoot, localWorkbenchAdapterAuthStore, normalizeWorkbenchAdapterAuthTarget, parseWorkbenchAdapterAuthTarget, sanitizeWorkbenchAdapterAuthBundle, } from "./adapter-auth.js";
@@ -26,14 +26,14 @@ export { createWorkbenchProgressStdoutParser, publishWorkbenchProgressStdoutEnve
 export { resolveSandboxTemplateImage, } from "./sandbox-backends/template-images.js";
 export { readOutputTraceFiles, workbenchTraceExecutionDirectory, workbenchTraceRunDirectory, workbenchTraceRunDirectoryName, } from "./trace-files.js";
 export { assertWorkbenchAdapterOperationSupport, assertWorkbenchAdapterOperationResultOk, collectWorkbenchAdapterOperationIssues, collectWorkbenchAdapterOperationRequirements, ensureWorkbenchAdapterOutputDir, WORKBENCH_ADAPTER_RESULT_FILE, normalizeWorkbenchAdapterOperationRequest, normalizeWorkbenchAdapterOperationResult, readWorkbenchAdapterOperationRequest, readWorkbenchAdapterOperationResult, workbenchAdapterOperationResultPath, writeWorkbenchAdapterOperationResult, } from "@workbench-ai/workbench-protocol";
-export { applyWorkbenchSubjectPatch, } from "./subject-patch.js";
+export { applyWorkbenchCandidatePatch, } from "./candidate-patch.js";
 export { createWorkbenchSandboxFileStore, createSandboxAdapterRequest, executionResultFromCompletedSandboxJob, materializeWorkbenchSandboxInput, readWorkbenchExecutionSpec, sanitizeWorkbenchExecutionJobForSandbox, } from "./sandbox-inputs.js";
 export { compileWorkbenchExecutionGraph, } from "./execution-graph.js";
-export { createBaselineSubjectExecution, createBaselineSubjectJob, createWorkbenchExecutionJob, expectedWorkbenchRunJobCount, engineCaseForCase, engineCaseIds, attemptJobCountForRunSpec, workbenchExecutionJobPurpose, MAX_WORKBENCH_RUN_BUDGET, planWorkbenchExecutionJobsForPurpose, validateWorkbenchRunEnvelope, workbenchExecutionJobId, } from "./execution-jobs.js";
+export { createBaselineCandidateExecution, createBaselineCandidateJob, createWorkbenchExecutionJob, expectedWorkbenchRunJobCount, engineCaseForCase, engineCaseIds, attemptJobCountForRunSpec, workbenchExecutionJobPurpose, MAX_WORKBENCH_RUN_BUDGET, planWorkbenchExecutionJobsForPurpose, validateWorkbenchRunEnvelope, workbenchExecutionJobId, } from "./execution-jobs.js";
 export { addCapacity, capacityFits, runWorkbenchExecutionDag, subtractCapacity, workbenchJobDependencies, workbenchJobHostCost, workbenchJobResources, } from "./execution-scheduler.js";
 export { assertWorkbenchExecutionIsolation, collectWorkbenchExecutionIsolationIssues, validateWorkbenchExecutionOutputPayloads, } from "./execution-outputs.js";
 export { collectSandboxAllocationScopeIssues, collectExecutionCapabilityScopeIssues, collectSandboxHandleScopeIssues, createWorkbenchSandboxAllocation, createWorkbenchSandboxExecutionMetadata, createWorkbenchExecutionCapability, executeValidatedSandboxExecution, } from "./sandbox-plane.js";
-export { buildSubjectCaseExecutionRefs, buildWorkbenchExecutionEvidence, isWorkbenchExecutionActive, readWorkbenchExecutionId, readWorkbenchExecutionMetadataNumber, readWorkbenchExecutionMetadataString, readWorkbenchExecutionPurpose, resolveWorkbenchJobGroupStatus, } from "./execution-evidence.js";
+export { buildCandidateCaseExecutionRefs, buildWorkbenchExecutionEvidence, isWorkbenchExecutionActive, readWorkbenchExecutionId, readWorkbenchExecutionMetadataNumber, readWorkbenchExecutionMetadataString, readWorkbenchExecutionPurpose, resolveWorkbenchJobGroupStatus, } from "./execution-evidence.js";
 export { buildWorkbenchTraceSessionsFromFiles, combineWorkbenchTraceSessions, finalizeWorkbenchExecutionTraceForJob, mergeWorkbenchExecutionTracesByJob, readWorkbenchExecutionTraceFiles, traceSessionLabel, } from "./execution-traces.js";
 export { DOCKER_SANDBOX_BACKEND, assertSandboxHostHealthForProvider, createDockerSandboxBackendDescriptor, createDockerSandboxPlane, resolveWorkbenchSandboxProviderName, sandboxProviderAdmissionForResources, sandboxProviderDefaultMaxConcurrentJobs, sandboxProviderLeaseScope, sandboxHostHealthExpectationForProvider, } from "./sandbox-backends/index.js";
 export const DEFAULT_ENVIRONMENT_VERSIONS = [
@@ -153,7 +153,7 @@ export const DEFAULT_ENVIRONMENTS = [
     {
         id: "env_node",
         name: "Node",
-        description: "Node runtime for JavaScript and TypeScript subjects.",
+        description: "Node runtime for JavaScript and TypeScript candidates.",
         currentVersionId: "envv_node_22",
         builtIn: true,
         createdAt: "2026-04-23T00:00:00.000Z",
@@ -191,8 +191,7 @@ function splitAuthoredSourceYaml(sourceYaml) {
     }
     const entries = [
         [BENCHMARK_SPEC_FILE, parsed.benchmark],
-        ["subjects/current/subject.yaml", splitSubjectSourceRecord(parsed.subject)],
-        ["optimizers/current.yaml", splitOptimizerSourceRecord(parsed.optimizer)],
+        ["candidates/current/candidate.yaml", splitCandidateSourceRecord(parsed.candidate)],
     ];
     return entries.flatMap(([filePath, value]) => {
         if (!value || typeof value !== "object" || Array.isArray(value)) {
@@ -204,23 +203,20 @@ function splitAuthoredSourceYaml(sourceYaml) {
             }];
     });
 }
-function splitSubjectSourceRecord(value) {
+function splitCandidateSourceRecord(value) {
     const record = cloneYamlRecord(value);
     if (!record) {
         return value;
     }
     delete record.benchmark;
     delete record.path;
-    rewriteAdapterSources(record, "subjects");
+    stripCandidateRuntimeSelection(record);
+    rewriteAdapterSources(record, "candidates/current");
     return record;
 }
-function splitOptimizerSourceRecord(value) {
-    const record = cloneYamlRecord(value);
-    if (!record) {
-        return value;
-    }
-    rewriteAdapterSources(record, "optimizers");
-    return record;
+function stripCandidateRuntimeSelection(record) {
+    delete record.selectedRunId;
+    delete record.selectedRunName;
 }
 function cloneYamlRecord(value) {
     return value && typeof value === "object" && !Array.isArray(value)
@@ -242,11 +238,10 @@ function sourcePathRelativeTo(yamlDir, sourcePath) {
 }
 function isAuthoredSourceYamlPath(filePath) {
     return filePath === BENCHMARK_SPEC_FILE ||
-        isWorkbenchSubjectManifestPath(filePath) ||
-        /^optimizers\/[^/]+\.ya?ml$/iu.test(filePath);
+        isWorkbenchCandidateManifestPath(filePath);
 }
-function formatOptimizerSummary(spec) {
-    return spec.improve ? `adapter:${spec.improve.use}` : "optimizer not configured";
+function formatImproveSummary(spec) {
+    return spec.improve ? `adapter:${spec.improve.use}` : "improve not configured";
 }
 function formatEngineRunSummary(spec) {
     return `adapter:${spec.engineRun.use}`;
@@ -287,10 +282,10 @@ function protocolStepForExecution(execution, manifests) {
     if (execution.purpose !== "improve") {
         throw new Error(`Protocol execution step only supports improve executions, not ${execution.purpose}.`);
     }
-    const operation = "optimizer.improve";
+    const operation = "candidate.improve";
     const command = adapterProtocolCommandSpec(execution.adapter, operation, manifests);
     return {
-        kind: "optimizer",
+        kind: "improver",
         label: execution.purpose,
         operation,
         executor: command.executor,
@@ -387,31 +382,31 @@ export function materializeWorkbenchRunResult(args) {
     const completed = args.jobs.filter((job) => job.status === "succeeded");
     const failedJobCount = args.jobs.filter((job) => job.status === "failed").length;
     const completedJobCount = args.jobs.filter((job) => job.status === "succeeded").length;
-    const subjectRevisions = completed
+    const candidateRevisions = completed
         .filter((job) => workbenchExecutionPurpose(job) === "improve")
-        .map((job) => normalizeSubjectRevisionJobOutput(job.output))
+        .map((job) => normalizeCandidateRevisionJobOutput(job.output))
         .filter((output) => output !== null)
         .sort((left, right) => left.attemptIndex - right.attemptIndex);
     const evaluationJobs = args.jobs.filter((job) => workbenchExecutionPurpose(job) === "attempt");
-    const evaluationsBySubject = new Map();
+    const evaluationsByCandidate = new Map();
     for (const job of evaluationJobs) {
-        const subjectId = readJobString(job.output, "subjectId") ??
-            readJobString(job.input, "subjectId") ??
-            job.subjectId;
-        if (subjectId) {
-            evaluationsBySubject.set(subjectId, [
-                ...(evaluationsBySubject.get(subjectId) ?? []),
+        const candidateId = readJobString(job.output, "candidateId") ??
+            readJobString(job.input, "candidateId") ??
+            job.candidateId;
+        if (candidateId) {
+            evaluationsByCandidate.set(candidateId, [
+                ...(evaluationsByCandidate.get(candidateId) ?? []),
                 job,
             ]);
         }
     }
-    const subjects = [];
-    const subjectFiles = {};
+    const candidates = [];
+    const candidateFiles = {};
     const evaluations = [];
-    for (const subjectRevision of subjectRevisions) {
-        const subjectId = subjectRevision.subjectId;
-        const subjectJobs = evaluationsBySubject.get(subjectId) ?? [];
-        const succeededEvaluationJobs = subjectJobs.filter((job) => job.status === "succeeded");
+    for (const candidateRevision of candidateRevisions) {
+        const candidateId = candidateRevision.candidateId;
+        const candidateJobs = evaluationsByCandidate.get(candidateId) ?? [];
+        const succeededEvaluationJobs = candidateJobs.filter((job) => job.status === "succeeded");
         const outputs = normalizeEvaluationSampleOutputs({
             jobs: succeededEvaluationJobs,
             allJobs: completed,
@@ -425,39 +420,38 @@ export function materializeWorkbenchRunResult(args) {
         ])
             .filter((key) => key !== null));
         const errorSampleJobs = [
-            ...subjectJobs.filter((job) => job.status === "failed"),
+            ...candidateJobs.filter((job) => job.status === "failed"),
             ...succeededEvaluationJobs.filter((job) => !outputJobIds.has(job.id)),
         ];
-        const errorSamples = errorEvaluationSamplesFromJobs(errorSampleJobs, subjectId, subjectRevision.attemptIndex, completedSampleKeys);
+        const errorSamples = errorEvaluationSamplesFromJobs(errorSampleJobs, candidateId, candidateRevision.attemptIndex, completedSampleKeys);
         const samples = [
             ...outputs.map(({ jobs, output }) => withJobUsage(output.sample, completed, jobs[0])),
             ...errorSamples,
         ].sort((left, right) => left.index - right.index || left.id.localeCompare(right.id));
-        const subjectName = normalizedSubjectDisplayName(args.spec.subject.name);
-        const evalRecord = createEvaluationRecord(subjectId, subjectName, samples);
+        const candidateName = normalizedCandidateDisplayName(args.spec.candidate.name);
+        const evalRecord = createEvaluationRecord(candidateId, candidateName, samples);
         const usage = mergeUsageSummaries([
-            subjectRevision.usage,
+            candidateRevision.usage,
             ...samples.map((sample) => sample.usage),
         ]);
-        const metrics = evaluationMeanMetrics(evalRecord);
-        const attemptIndex = subjectRevision.attemptIndex;
+        const attemptIndex = candidateRevision.attemptIndex;
         const evaluationTraces = [
             ...outputs.flatMap(({ output }) => output.traces),
             ...errorSampleJobs.flatMap(jobTracePaths),
         ].sort();
-        const baseId = subjectRevision.baseId && subjectRevision.baseId !== subjectId
-            ? subjectRevision.baseId
+        const baseId = candidateRevision.baseId && candidateRevision.baseId !== candidateId
+            ? candidateRevision.baseId
             : null;
-        const sourceMeta = subjectSourceMetadata(args.subjectSourceFiles);
+        const sourceMeta = candidateSourceMetadata(args.candidateSourceFiles);
         const benchmarkMeta = benchmarkSourceMetadata(args.benchmarkSourceFiles);
         const meta = {
             attemptIndex,
             sampleCount: evalRecord.sampleCount,
-            optimizer: formatOptimizerSummary(args.spec),
+            improver: formatImproveSummary(args.spec),
             engineRun: formatEngineRunSummary(args.spec),
             strategy: "greedy",
             traces: {
-                improve: subjectRevision.traces,
+                improve: candidateRevision.traces,
                 evaluations: evaluationTraces,
             },
         };
@@ -467,52 +461,114 @@ export function materializeWorkbenchRunResult(args) {
         if (benchmarkMeta) {
             meta.benchmark = benchmarkMeta;
         }
-        const record = {
-            id: subjectId,
-            ...(subjectName ? { name: subjectName } : {}),
-            ordinal: args.existingSubjectCount + subjects.length,
-            benchmarkFingerprint: args.benchmarkFingerprint,
-            subjectFingerprint: args.subjectFingerprint ?? materializedSubjectFingerprint(args.spec, subjectRevision.files),
-            createdAt: args.startedAt,
-            ...(baseId ? { baseId } : {}),
-            referenceIds: [],
-            status: evalRecord.completedSampleCount > 0 ? "evaluated" : "eval_error",
-            fileChanges: subjectRevision.fileChanges,
-            ...(metrics ? { metrics } : {}),
-            ...(usage ? { usage } : {}),
-            eval: evalRecord,
-            ...(subjectRevision.prompt ? { prompt: subjectRevision.prompt } : {}),
-            meta,
-        };
-        subjects.push(record);
+        const record = preserveExistingCandidateIdentity({
+            candidate: {
+                id: candidateId,
+                ...(candidateName ? { name: candidateName } : {}),
+                version: args.existingCandidateCount + candidates.length + 1,
+                ordinal: args.existingCandidateCount + candidates.length + 1,
+                benchmarkFingerprint: args.benchmarkFingerprint,
+                candidateFingerprint: args.candidateFingerprint ?? materializedCandidateFingerprint(args.spec, candidateRevision.files),
+                createdAt: args.startedAt,
+                ...(baseId ? { baseId } : {}),
+                referenceIds: [],
+                status: evalRecord.completedSampleCount > 0 ? "evaluated" : "eval_error",
+                fileChanges: candidateRevision.fileChanges,
+                ...(usage ? { usage } : {}),
+                eval: evalRecord,
+                ...(candidateRevision.prompt ? { prompt: candidateRevision.prompt } : {}),
+                meta,
+            },
+            previousCandidate: args.previousCandidate ?? null,
+        });
+        candidates.push(record);
         evaluations.push(createEvaluationScorecard({
             runId: args.runId,
             benchmarkFingerprint: args.benchmarkFingerprint,
             createdAt: args.startedAt,
-            subject: record,
+            candidate: record,
+            candidateRunId: args.spec.candidate.selectedRunId,
+            candidateRunName: args.spec.candidate.selectedRunName,
             evaluation: evalRecord,
         }));
-        subjectFiles[subjectId] = materializedSubjectFiles({
-            subjectRevisionFiles: subjectRevision.files,
+        candidateFiles[candidateId] = materializedCandidateFiles({
+            candidateRevisionFiles: candidateRevision.files,
         });
     }
-    const selectedSubject = selectSubject({
-        subjects,
-        previousSubject: args.previousSubject ?? null,
+    const selectedCandidate = selectCandidate({
+        candidates,
+        previousCandidate: args.previousCandidate ?? null,
     });
     return {
-        subjects,
-        subjectFiles,
+        candidates,
+        candidateFiles,
         evaluations,
-        activeSubjectId: selectedSubject?.id ?? args.previousSubject?.id ?? null,
-        selectedSubject,
+        activeCandidateId: selectedCandidate?.id ?? args.previousCandidate?.id ?? null,
+        selectedCandidate,
         completedJobCount,
         failedJobCount,
     };
 }
-function subjectSourceMetadata(files) {
+function preserveExistingCandidateIdentity(args) {
+    const previous = args.previousCandidate;
+    if (!previous || previous.id !== args.candidate.id) {
+        return args.candidate;
+    }
+    const baseId = args.candidate.baseId ?? previous.baseId;
+    const prompt = args.candidate.prompt ?? previous.prompt;
+    const meta = mergeExistingCandidateMeta(previous.meta, args.candidate.meta);
+    return {
+        ...args.candidate,
+        version: previous.version,
+        ordinal: previous.version,
+        createdAt: previous.createdAt,
+        ...(args.candidate.name ?? previous.name
+            ? { name: (args.candidate.name ?? previous.name) }
+            : {}),
+        ...(baseId ? { baseId } : {}),
+        referenceIds: previous.referenceIds.length > 0
+            ? [...previous.referenceIds]
+            : args.candidate.referenceIds,
+        fileChanges: args.candidate.fileChanges.length > 0
+            ? args.candidate.fileChanges
+            : [...previous.fileChanges],
+        ...(prompt ? { prompt } : {}),
+        ...(meta ? { meta } : {}),
+    };
+}
+function mergeExistingCandidateMeta(previousMeta, candidateMeta) {
+    const previous = jsonRecord(previousMeta);
+    const candidate = jsonRecord(candidateMeta);
+    if (!previous) {
+        return candidateMeta;
+    }
+    if (!candidate) {
+        return previousMeta;
+    }
+    const previousTraces = jsonRecord(previous.traces);
+    const candidateTraces = jsonRecord(candidate.traces);
+    if (!previousTraces || !candidateTraces) {
+        return { ...previous, ...candidate };
+    }
+    const traces = {
+        ...previousTraces,
+        ...candidateTraces,
+    };
+    const candidateImproveTraces = Array.isArray(candidateTraces.improve)
+        ? candidateTraces.improve
+        : [];
+    if (candidateImproveTraces.length === 0 && previousTraces.improve !== undefined) {
+        traces.improve = previousTraces.improve;
+    }
+    return {
+        ...previous,
+        ...candidate,
+        traces,
+    };
+}
+function candidateSourceMetadata(files) {
     const sourceFiles = (files ?? [])
-        .filter((file) => /^subjects\/[^/]+\/subject\.ya?ml$/iu.test(file.path))
+        .filter((file) => /^candidates\/[^/]+\/candidate\.ya?ml$/iu.test(file.path))
         .sort((left, right) => left.path.localeCompare(right.path))
         .map((file) => ({
         path: file.path,
@@ -536,14 +592,13 @@ function benchmarkSourceMetadata(files) {
     }));
     return sourceFiles.length > 0 ? { files: sourceFiles } : null;
 }
-function materializedSubjectFingerprint(spec, files) {
+function materializedCandidateFingerprint(spec, files) {
     const hash = createHash("sha256");
-    hash.update("workbench-subject-v1\0");
-    hash.update("materialized\0runner\0");
-    hash.update(JSON.stringify(spec.run));
+    hash.update("workbench-candidate-v1\0");
+    hash.update("materialized\0");
     hash.update("prepare");
-    hash.update(JSON.stringify(spec.subject.prepare ?? null));
-    for (const file of filterSubjectSourceFiles(files).slice().sort((left, right) => left.path.localeCompare(right.path))) {
+    hash.update(JSON.stringify(spec.candidate.prepare ?? null));
+    for (const file of filterCandidateSourceFiles(files).slice().sort((left, right) => left.path.localeCompare(right.path))) {
         hash.update("\0file\0");
         hash.update(file.path);
         hash.update("\0");
@@ -555,9 +610,9 @@ function materializedSubjectFingerprint(spec, files) {
     }
     return hash.digest("hex");
 }
-function materializedSubjectFiles(args) {
+function materializedCandidateFiles(args) {
     const byPath = new Map();
-    for (const file of filterSubjectSourceFiles(args.subjectRevisionFiles)) {
+    for (const file of filterCandidateSourceFiles(args.candidateRevisionFiles)) {
         byPath.set(file.path, { ...file });
     }
     return [...byPath.values()].sort((left, right) => left.path.localeCompare(right.path));
@@ -565,12 +620,15 @@ function materializedSubjectFiles(args) {
 function createEvaluationScorecard(args) {
     const evaluation = args.evaluation;
     return {
-        id: evaluationScorecardId(args.runId, args.subject.id),
+        id: evaluationScorecardId(args.runId, args.candidate.id),
         runId: args.runId,
         benchmarkFingerprint: args.benchmarkFingerprint,
-        subjectFingerprint: args.subject.subjectFingerprint,
-        subjectId: args.subject.id,
-        ...(args.subject.name ? { subjectName: args.subject.name } : {}),
+        candidateFingerprint: args.candidate.candidateFingerprint,
+        candidateId: args.candidate.id,
+        ...(args.candidate.name ? { candidateName: args.candidate.name } : {}),
+        candidateVersion: args.candidate.version,
+        ...(args.candidateRunId ? { candidateRunId: args.candidateRunId } : {}),
+        ...(args.candidateRunName ? { candidateRunName: args.candidateRunName } : {}),
         createdAt: args.createdAt,
         updatedAt: evaluation.finishedAt ?? args.createdAt,
         status: evaluation.status,
@@ -584,10 +642,10 @@ function createEvaluationScorecard(args) {
         evaluation,
     };
 }
-export function evaluationScorecardId(runId, subjectId) {
+export function evaluationScorecardId(runId, candidateId) {
     const runPart = runId.replace(/[^a-z0-9]+/giu, "_").replace(/^_+|_+$/gu, "").slice(-24);
-    const subjectPart = subjectId.replace(/[^a-z0-9]+/giu, "_").replace(/^_+|_+$/gu, "").slice(-24);
-    return `eval_${runPart}_${subjectPart}`;
+    const candidatePart = candidateId.replace(/[^a-z0-9]+/giu, "_").replace(/^_+|_+$/gu, "").slice(-24);
+    return `eval_${runPart}_${candidatePart}`;
 }
 export function selectExecutionOutputFilesForInspection(args) {
     return args.files.filter((file) => !isWorkbenchInternalOutputPath(file.path));
@@ -602,56 +660,89 @@ export function isWorkbenchInternalOutputPath(filePath) {
         normalized === "exit_code" ||
         /^[a-z_-]+_(stdout\.log|stderr\.log|exit_code)$/u.test(normalized));
 }
-export function createSubjectRevisionTraceInputFiles(args) {
+export function createOptimizerTraceInputFiles(args) {
     const files = [];
-    const manifestJobs = [];
+    const executions = [];
     const jobs = args.jobs
-        .filter((job) => job.runId === args.runId && isTerminalExecutionJob(job))
+        .filter(isOptimizerTraceInputJob)
         .sort(compareTraceInputJobs);
-    for (const job of jobs) {
+    jobs.forEach((job, index) => {
+        const sequence = String(index + 1).padStart(6, "0");
+        const executionPath = `executions/${sequence}`;
+        const operation = "engine.run";
         const jobFiles = completedJobOutputFiles(job);
-        const rawTraceFiles = jobFiles.filter((file) => normalizeRelativePath(file.path).startsWith(".workbench/traces/"));
-        files.push(...rawTraceFiles.map((file) => ({ ...file })));
-        const events = args.events
-            .filter((event) => event.runId === args.runId && event.jobId === job.id)
-            .sort((left, right) => left.at.localeCompare(right.at));
-        const eventPath = `events/${job.id}.ndjson`;
-        if (events.length > 0) {
-            files.push(textSurfaceFile(eventPath, `${events.map((event) => JSON.stringify(event)).join("\n")}\n`));
-        }
-        const summaryPath = `jobs/${job.id}.json`;
-        const summary = subjectRevisionTraceJobSummary(job, {
-            eventPath: events.length > 0 ? eventPath : null,
-            rawTracePaths: rawTraceFiles.map((file) => file.path).sort(),
-        });
-        files.push(textSurfaceFile(summaryPath, `${JSON.stringify(summary, null, 2)}\n`));
-        manifestJobs.push({
-            ...summary,
-            summary_path: summaryPath,
+        const requestFile = traceInputRequestFile(jobFiles, operation);
+        const resultFile = traceInputResultFile(jobFiles, operation);
+        const requestPath = `${executionPath}/request.json`;
+        const resultPath = `${executionPath}/result.json`;
+        const filesPath = `${executionPath}/files`;
+        files.push(textSurfaceFile(requestPath, requestFile?.content ?? `${JSON.stringify(traceInputRequestFallback(job, operation), null, 2)}\n`));
+        files.push(textSurfaceFile(resultPath, resultFile?.content ?? `${JSON.stringify(traceInputResultFallback(job, operation), null, 2)}\n`));
+        files.push(...jobFiles.map((file) => ({
+            ...file,
+            path: normalizeRelativePath(`${filesPath}/${file.path}`),
+        })));
+        executions.push({
+            path: executionPath,
+            operation,
+            status: job.status,
+            candidateId: job.candidateId ?? readJobString(job.input, "candidateId") ?? null,
+            runId: job.runId,
+            jobId: job.id,
+            attemptIndex: readOptionalJobNumber(job.input, "attemptIndex") ?? null,
+            sampleIndex: readOptionalJobNumber(job.input, "sampleIndex") ?? null,
+            caseId: readJobString(job.input, "caseId") ?? null,
+            requestPath,
+            resultPath,
+            filesPath,
         });
-    }
-    files.push(textSurfaceFile("manifest.json", `${JSON.stringify({
-        run_id: args.runId,
-        jobs: manifestJobs,
+    });
+    files.push(textSurfaceFile("index.json", `${JSON.stringify({
+        schema: "workbench.optimizer-traces.v1",
+        executions,
     }, null, 2)}\n`));
     return dedupeSurfaceFiles(files);
 }
-export function createSubjectEvaluationTraceInputFiles(args) {
-    const subject = args.subject;
-    if (!subject?.eval && !subject?.metrics) {
-        return [];
+export function evaluationMeanMetrics(evaluation) {
+    const entries = Object.entries(evaluation?.metrics ?? {})
+        .filter((entry) => Number.isFinite(entry[1].mean));
+    return entries.length > 0
+        ? Object.fromEntries(entries.map(([key, stats]) => [key, stats.mean]))
+        : undefined;
+}
+export function candidateRecordWithoutDerivedFields(candidate) {
+    const { metrics: _metrics, candidateRunId: _candidateRunId, candidateRunName: _candidateRunName, ...record } = candidate;
+    return record;
+}
+export function candidateSummaryFromRecord(candidate) {
+    const { eval: _eval, prompt: _prompt, meta: _meta, ...summary } = candidateRecordWithoutDerivedFields(candidate);
+    return summary;
+}
+export function workbenchRunExecutionFingerprint(args) {
+    const hash = createHash("sha256");
+    hash.update("workbench-run-execution-v1\0");
+    hash.update(args.specVersionId ?? "");
+    hash.update("\0");
+    hash.update(args.environmentVersionId ?? "");
+    hash.update("\0");
+    hash.update(args.sourceYaml ?? "");
+    for (const file of (args.adapterFiles ?? []).slice().sort((left, right) => left.path.localeCompare(right.path))) {
+        hash.update("\0file\0");
+        hash.update(file.path);
+        hash.update("\0");
+        hash.update(file.kind);
+        hash.update("\0");
+        hash.update(file.encoding);
+        hash.update("\0");
+        hash.update(file.executable ? "1" : "0");
+        hash.update("\0");
+        hash.update(file.content);
     }
-    const filePath = normalizeRelativePath(args.path ?? `base-subject/${subject.id}/evaluation.json`);
-    const payload = {
-        kind: "subject_evaluation",
-        subjectId: subject.id,
-        status: subject.status,
-        metrics: subject.metrics ?? null,
-        fileChanges: subject.fileChanges,
-        eval: subject.eval ?? null,
-        prompt: subject.prompt ?? null,
-    };
-    return [textSurfaceFile(filePath, `${JSON.stringify(payload, null, 2)}\n`)];
+    return hash.digest("hex");
+}
+function isOptimizerTraceInputJob(job) {
+    return isTerminalExecutionJob(job) &&
+        workbenchExecutionPurpose(job) === "attempt";
 }
 function isTerminalExecutionJob(job) {
     return job.kind === "execute" && (job.status === "succeeded" ||
@@ -662,20 +753,10 @@ function compareTraceInputJobs(left, right) {
     const leftAttempt = readOptionalJobNumber(left.input, "attemptIndex") ?? -1;
     const rightAttempt = readOptionalJobNumber(right.input, "attemptIndex") ?? -1;
     return leftAttempt - rightAttempt ||
-        purposeSortKey(workbenchExecutionPurpose(left)) - purposeSortKey(workbenchExecutionPurpose(right)) ||
         (readOptionalJobNumber(left.input, "sampleIndex") ?? -1) - (readOptionalJobNumber(right.input, "sampleIndex") ?? -1) ||
         (readJobString(left.input, "caseId") ?? "").localeCompare(readJobString(right.input, "caseId") ?? "") ||
         left.id.localeCompare(right.id);
 }
-function purposeSortKey(purpose) {
-    if (purpose === "improve") {
-        return 0;
-    }
-    if (purpose === "attempt") {
-        return 1;
-    }
-    return 3;
-}
 function completedJobOutputFiles(job) {
     const output = jsonRecord(job.output);
     if (!Array.isArray(output.files)) {
@@ -689,35 +770,70 @@ function completedJobOutputFiles(job) {
     }
     return files;
 }
-function subjectRevisionTraceJobSummary(job, paths) {
-    const output = jsonRecord(job.output);
+function traceInputRequestFile(files, operation) {
+    return files.find((file) => {
+        const normalized = normalizeRelativePath(file.path);
+        return normalized.startsWith(".workbench/traces/") &&
+            normalized.endsWith("/request.json") &&
+            file.encoding === "utf8" &&
+            traceJsonOperation(file) === operation;
+    }) ?? null;
+}
+function traceInputResultFile(files, operation) {
+    return files.find((file) => {
+        const normalized = normalizeRelativePath(file.path);
+        return normalized.startsWith(".workbench/traces/") &&
+            normalized.endsWith("/result.json") &&
+            file.encoding === "utf8" &&
+            traceJsonOperation(file) === operation;
+    }) ?? null;
+}
+function traceJsonOperation(file) {
+    try {
+        const parsed = JSON.parse(file.content);
+        return typeof parsed?.operation === "string" ? parsed.operation : null;
+    }
+    catch {
+        return null;
+    }
+}
+function traceInputRequestFallback(job, operation) {
+    const execution = jsonRecord(jsonRecord(job.input).execution);
     return {
-        job_id: job.id,
-        purpose: workbenchExecutionPurpose(job) ?? "unknown",
-        status: job.status,
-        subject_id: job.subjectId ?? readJobString(job.input, "subjectId"),
-        attempt_index: readOptionalJobNumber(job.input, "attemptIndex"),
-        sample_index: readOptionalJobNumber(job.input, "sampleIndex"),
-        case_id: readJobString(job.input, "caseId"),
-        created_at: job.createdAt,
-        ...(job.startedAt ? { started_at: job.startedAt } : {}),
-        ...(job.finishedAt ? { finished_at: job.finishedAt } : {}),
-        ...(job.error ? { error: job.error } : {}),
-        traces: jobTracePaths(job),
-        event_path: paths.eventPath,
-        raw_trace_paths: [...paths.rawTracePaths],
-        output: summarizeJobOutputForTrace(output),
+        protocol: "workbench.adapter.v3",
+        id: typeof execution.id === "string" ? execution.id : job.id,
+        jobId: job.id,
+        operation,
+        invocation: jsonRecord(execution.adapter),
+        context: {
+            candidate: {
+                id: job.candidateId ?? readJobString(job.input, "candidateId") ?? null,
+            },
+            attempt: {
+                attemptIndex: readOptionalJobNumber(job.input, "attemptIndex") ?? null,
+                sampleIndex: readOptionalJobNumber(job.input, "sampleIndex") ?? null,
+                caseId: readJobString(job.input, "caseId") ?? null,
+            },
+        },
     };
 }
-function summarizeJobOutputForTrace(output) {
-    const { files: _files, fileSet: _fileSet, subjectPatch, ...rest } = output;
-    const patch = jsonRecord(subjectPatch);
-    const { files: _patchFiles, ...patchSummary } = patch;
+function traceInputResultFallback(job, operation) {
+    const output = jsonRecord(job.output);
+    const ok = job.status === "succeeded" && output.ok !== false;
+    const value = operation === "candidate.improve"
+        ? jsonRecord(output.candidatePatch)
+        : operation === "engine.run"
+            ? jsonRecord(output.result)
+            : {};
     return {
-        ...rest,
-        ...(Object.keys(patch).length > 0
-            ? { subjectPatch: patchSummary }
-            : {}),
+        protocol: "workbench.adapter-result.v1",
+        operation,
+        ok,
+        ...(Object.keys(value).length > 0 ? { value: value } : {}),
+        ...(typeof output.summary === "string" ? { summary: output.summary } : {}),
+        ...(output.feedback !== undefined ? { feedback: output.feedback } : {}),
+        ...(output.usage !== undefined ? { usage: output.usage } : {}),
+        ...(!ok ? { error: job.error ?? "Execution did not complete successfully." } : {}),
     };
 }
 function textSurfaceFile(path, content) {
@@ -744,7 +860,7 @@ export function buildWorkbenchProjectSourceFiles(input) {
         ...(input.specFiles
             ? input.specFiles.map((file) => ({ ...file }))
             : [textSurfaceFile("benchmark.yaml", input.specSource ?? "")]),
-        ...prefixProjectSourceFiles(input.subjectFiles, input.subjectFilesPath),
+        ...prefixProjectSourceFiles(input.candidateFiles, input.candidateFilesPath),
         ...prefixProjectSourceFiles(input.engineResolveFiles, input.engineResolveFilesPath),
         ...(input.adapterFiles ?? []).map((file) => ({ ...file })),
         ...(input.dockerfiles ?? []).map((file) => ({ ...file })),
@@ -772,18 +888,18 @@ function prefixProjectSourceFiles(files, rootPath) {
         };
     });
 }
-export function isSubjectSourceFilePath(filePath) {
+export function isCandidateSourceFilePath(filePath) {
     const normalized = normalizeRelativePath(filePath);
     return (normalized !== ".workbench" &&
         !normalized.startsWith(".workbench/") &&
         normalized !== "workbench-result.json");
 }
-export function filterSubjectSourceFiles(files) {
+export function filterCandidateSourceFiles(files) {
     return files
-        .filter((file) => isSubjectSourceFilePath(file.path))
+        .filter((file) => isCandidateSourceFilePath(file.path))
         .map((file) => ({ ...file }));
 }
-export function buildSubjectLineage(args) {
+export function buildCandidateLineage(args) {
     const orderedSummaries = args.summaries.slice().sort((left, right) => {
         const createdAt = left.createdAt.localeCompare(right.createdAt);
         return createdAt !== 0 ? createdAt : left.id.localeCompare(right.id);
@@ -856,7 +972,7 @@ function globPatternToRegExp(pattern) {
 function escapeRegExp(value) {
     return value.replace(/[\\^$.*+?()[\]{}|]/gu, "\\$&");
 }
-export function summarizeSubjectFiles(files, changedPaths = files.map((file) => file.path)) {
+export function summarizeCandidateFiles(files, changedPaths = files.map((file) => file.path)) {
     const changed = new Set(changedPaths);
     return [...files]
         .sort((left, right) => left.path.localeCompare(right.path))
@@ -875,7 +991,7 @@ export function summarizeSubjectFiles(files, changedPaths = files.map((file) =>
         };
     });
 }
-export function createSubjectFilePreview(args) {
+export function createCandidateFilePreview(args) {
     if (args.view === "diff") {
         throw new Error("Diff previews require explicit before and after file content.");
     }
@@ -901,14 +1017,14 @@ export function createSubjectFilePreview(args) {
 export function createCaseReview(args) {
     const preferredSampleIndex = uniqueExecutionSampleIndex(args.executions ?? []);
     const sampleMatchesCase = (sample) => (sample.cases ?? []).some((entry) => entry.id === args.caseId);
-    const samples = args.subject.eval?.samples ?? [];
+    const samples = args.candidate.eval?.samples ?? [];
     const sampleResult = samples.find((sample) => typeof preferredSampleIndex === "number" &&
         sample.index === preferredSampleIndex &&
         sampleMatchesCase(sample)) ?? samples.find(sampleMatchesCase);
     const caseResult = sampleResult?.cases?.find((entry) => entry.id === args.caseId);
     if (!sampleResult && (args.executions?.length ?? 0) > 0) {
         return {
-            subjectId: args.subject.id,
+            candidateId: args.candidate.id,
             caseId: args.caseId,
             caseLabel: args.caseId,
             ...(typeof preferredSampleIndex === "number"
@@ -920,13 +1036,13 @@ export function createCaseReview(args) {
         };
     }
     if (!sampleResult) {
-        throw new Error(`Case ${args.caseId} was not found on subject ${args.subject.id}.`);
+        throw new Error(`Case ${args.caseId} was not found on candidate ${args.candidate.id}.`);
     }
     const durationMs = typeof caseResult?.durationMs === "number"
         ? caseResult.durationMs
         : undefined;
     return {
-        subjectId: args.subject.id,
+        candidateId: args.candidate.id,
         caseId: caseResult?.id ?? args.caseId,
         caseLabel: caseResult?.label ?? args.caseId,
         sampleId: sampleResult.id,
@@ -965,37 +1081,39 @@ function parseAuthoredWorkbenchSourceSpec(source) {
     }
     const resolved = resolveWorkbenchResolvedSourceYamlInternal(source);
     return {
-        version: 3,
+        version: 4,
         benchmark: {
             name: resolved.benchmark.name,
             description: resolved.benchmark.description,
             engine: authoredAdapterSpecFromInvocation(resolved.engine),
         },
-        subject: {
-            name: resolved.subject.name,
-            description: resolved.subject.description,
-            files: { path: resolved.subject.files.path },
-            ...(resolved.subject.prepare ? { prepare: { ...resolved.subject.prepare } } : {}),
-            run: runSpecFromInvocation(resolved.run),
-        },
-        ...(resolved.optimizer
-            ? {
-                optimizer: {
-                    name: resolved.optimizer.name,
-                    ...(resolved.optimizer.description ? { description: resolved.optimizer.description } : {}),
-                    edits: [...resolved.optimizer.edits],
-                    improve: improveSpecFromInvocation(resolved.improve),
+        candidate: {
+            name: resolved.candidate.name,
+            description: resolved.candidate.description,
+            files: { path: resolved.candidate.files.path },
+            ...(resolved.candidate.prepare ? { prepare: { ...resolved.candidate.prepare } } : {}),
+            defaultRun: resolved.candidate.defaultRun,
+            runs: Object.fromEntries(Object.entries(resolved.candidate.runs).map(([runId, run]) => [
+                runId,
+                {
+                    name: run.name,
+                    ...authoredAdapterSpecFromInvocation(run),
                 },
-            }
-            : {}),
+            ])),
+            ...(resolved.candidate.improve
+                ? {
+                    improve: {
+                        edits: [...resolved.candidate.improve.edits],
+                        ...improveSpecFromInvocation(resolved.improve),
+                    },
+                }
+                : {}),
+        },
     };
 }
 function improveSpecFromInvocation(invocation) {
     return authoredAdapterSpecFromInvocation(invocation);
 }
-function runSpecFromInvocation(invocation) {
-    return authoredAdapterSpecFromInvocation(invocation);
-}
 function authoredAdapterSpecFromInvocation(invocation) {
     const config = jsonRecord(invocation.with);
     return {
@@ -1048,9 +1166,9 @@ export function createWorkbenchRunWorkload(args) {
     if (!purpose) {
         throw new Error(`Unsupported runtime job kind: ${args.job.kind}`);
     }
-    const subjectId = readJobString(args.job.input, "subjectId") ?? args.job.subjectId;
-    if (!subjectId) {
-        throw new Error(`${purpose} execution job is missing subjectId.`);
+    const candidateId = readJobString(args.job.input, "candidateId") ?? args.job.candidateId;
+    if (!candidateId) {
+        throw new Error(`${purpose} execution job is missing candidateId.`);
     }
     const attemptIndex = readRequiredJobNumber(args.job.input, "attemptIndex", `${purpose} execution job`);
     const sampleIndex = purpose === "improve"
@@ -1066,7 +1184,7 @@ export function createWorkbenchRunWorkload(args) {
         ? engineCaseFilesForRuntimeInput({ spec: args.spec, engineCase })
         : [];
     const engineCaseSpec = engineCase?.case;
-    const initial = createInitialSubjectFiles({
+    const initial = createInitialCandidateFiles({
         baseFiles: args.baseFiles,
         spec: args.spec,
         attemptIndex,
@@ -1074,10 +1192,10 @@ export function createWorkbenchRunWorkload(args) {
     return {
         job: args.job,
         spec: args.spec,
-        subjectId,
+        candidateId,
         attemptIndex,
         sampleIndex,
-        subjectFiles: initial.files,
+        candidateFiles: initial.files,
         caseId,
         engineResolveFiles: selectedEngineResolveFiles,
         traceFiles: (args.traceFiles ?? []).map((file) => ({ ...file })),
@@ -1088,22 +1206,22 @@ export function createWorkbenchRunWorkload(args) {
         baseId: readJobString(args.job.input, "baseId"),
     };
 }
-function createInitialSubjectFiles(args) {
-    const editablePaths = optimizerEdits(args.spec).map(normalizeRelativePath);
+function createInitialCandidateFiles(args) {
+    const editablePaths = improveEdits(args.spec).map(normalizeRelativePath);
     const editPath = editablePaths[0];
-    const subjectPaths = editPath ? [editPath] : [];
+    const candidatePaths = editPath ? [editPath] : [];
     const files = args.baseFiles.length > 0
         ? args.baseFiles.map((file) => ({ ...file }))
         : editPath
             ? normalizeSurfaceFiles([{ path: editPath, content: "" }])
             : [];
     const prompt = [
-        `Run the subject workload for benchmark: ${args.spec.benchmark.description}`,
-        `Attempt ${args.attemptIndex + 1} uses ${formatOptimizerSummary(args.spec)}; the improve adapter may edit the subject before Workbench scores it.`,
+        `Run the candidate workload for benchmark: ${args.spec.benchmark.description}`,
+        `Attempt ${args.attemptIndex + 1} uses ${formatImproveSummary(args.spec)}; the improve adapter may edit the candidate before Workbench scores it.`,
     ].join("\n");
     const byPath = new Map(files.map((file) => [file.path, file]));
     if (editPath &&
-        ![...byPath.keys()].some((filePath) => subjectPaths.includes(filePath))) {
+        ![...byPath.keys()].some((filePath) => candidatePaths.includes(filePath))) {
         byPath.set(editPath, {
             path: editPath,
             kind: "text",
@@ -1167,7 +1285,7 @@ export function workbenchExecutionExecutorForRuntimeInput(args) {
 }
 function adapterOperationForExecutionPurpose(purpose) {
     if (purpose === "improve") {
-        return "optimizer.improve";
+        return "candidate.improve";
     }
     if (purpose === "attempt") {
         return "engine.run";
@@ -1281,8 +1399,8 @@ function normalizeRuntimeControlInputs(value) {
     }
     const record = value;
     const inputs = {};
-    if (hasOwn(record, "subject")) {
-        inputs.subject = normalizeRuntimeControlFiles(record.subject, "inputs.subject");
+    if (hasOwn(record, "candidate")) {
+        inputs.candidate = normalizeRuntimeControlFiles(record.candidate, "inputs.candidate");
     }
     if (hasOwn(record, "case")) {
         inputs.case = normalizeRuntimeControlFiles(record.case, "inputs.case");
@@ -1326,8 +1444,8 @@ function normalizeRuntimeControlOperation(value, label) {
     const operation = record.operation;
     if (operation !== "engine.resolve" &&
         operation !== "engine.run" &&
-        operation !== "subject.run" &&
-        operation !== "optimizer.improve") {
+        operation !== "candidate.run" &&
+        operation !== "candidate.improve") {
         throw new Error(`Workbench runtime-control ${label}.operation is invalid.`);
     }
     const invocation = record.invocation;
@@ -1415,7 +1533,7 @@ export async function executeAdapterInCurrentRuntime(args, execution, startedAt,
     };
     try {
         if (execution.purpose === "improve") {
-            return await executeSubjectRevisionExecutionInCurrentRuntime(runtimeInput, execution, startedAt, capability, eventPublisher);
+            return await executeCandidateRevisionExecutionInCurrentRuntime(runtimeInput, execution, startedAt, capability, eventPublisher);
         }
         if (execution.purpose === "attempt") {
             return await executeAttemptExecutionInCurrentRuntime(runtimeInput, execution, startedAt, capability, eventPublisher);
@@ -1589,22 +1707,22 @@ function completedJobFromSandboxResult(fallbackJob, startedAt, result) {
     }
     return attachSandboxMetadataToJob(failWorkbenchRunJob(fallbackJob, result.startedAt || startedAt, result.error ?? `Sandbox execution ${result.status}.`, result.finishedAt), asRuntimeRecord(result.metadata).sandbox);
 }
-async function executeSubjectRevisionExecutionInCurrentRuntime(args, execution, startedAt, capability, eventPublisher) {
+async function executeCandidateRevisionExecutionInCurrentRuntime(args, execution, startedAt, capability, eventPublisher) {
     const { workload, result } = await runHostedProtocolExecutionResult(args, execution, startedAt, capability, eventPublisher);
     if (result.error || (result.exitCode ?? 0) !== 0) {
         return failWorkbenchRunJob(args.job, startedAt, result.error ?? `Adapter ${execution.adapter.use} exited with status ${result.exitCode}.`, result.finishedAt, result);
     }
     const finishedAt = result.finishedAt ?? new Date().toISOString();
-    const subjectPatch = createSubjectPatchFromResult(result, args.spec);
-    if (subjectPatch.fileChanges.length === 0) {
-        return failWorkbenchRunJob(args.job, startedAt, `${execution.adapter.use === "command" ? "Command improve adapter" : `Adapter ${execution.adapter.use}`} completed without changing subject files covered by optimizer edits.`, finishedAt, result);
-    }
-    const subjectRevisionFiles = applyWorkbenchSubjectPatch({
-        baseFiles: workload.subjectFiles,
-        patch: subjectPatch,
-        edits: requireOptimizerEdits(args.spec),
+    const candidatePatch = createCandidatePatchFromResult(result, args.spec);
+    if (candidatePatch.fileChanges.length === 0) {
+        return failWorkbenchRunJob(args.job, startedAt, `${execution.adapter.use === "command" ? "Command improve adapter" : `Adapter ${execution.adapter.use}`} completed without changing candidate files covered by improve edits.`, finishedAt, result);
+    }
+    const candidateRevisionFiles = applyWorkbenchCandidatePatch({
+        baseFiles: workload.candidateFiles,
+        patch: candidatePatch,
+        edits: requireImproveEdits(args.spec),
     });
-    const usage = assignUsageRole("optimizer", result.usage);
+    const usage = assignUsageRole("improver", result.usage);
     return {
         ...args.job,
         status: "succeeded",
@@ -1616,13 +1734,13 @@ async function executeSubjectRevisionExecutionInCurrentRuntime(args, execution,
             ok: true,
             executionId: execution.id,
             purpose: execution.purpose,
-            subjectId: workload.subjectId,
+            candidateId: workload.candidateId,
             attemptIndex: workload.attemptIndex,
             baseId: workload.baseId,
             prompt: workload.prompt,
-            subjectPatch,
-            fileChanges: subjectPatch.fileChanges,
-            files: subjectRevisionFiles,
+            candidatePatch,
+            fileChanges: candidatePatch.fileChanges,
+            files: candidateRevisionFiles,
             traces: traceFilePaths(result.files),
             ...(usage ? { usage } : {}),
             ...(result.summary !== undefined ? { summary: result.summary } : {}),
@@ -1655,7 +1773,7 @@ async function executeAttemptExecutionInCurrentRuntime(args, execution, startedA
     const finishedAt = workloadResult.finishedAt ?? new Date().toISOString();
     const usage = attemptUsageSummary(workloadResult.usage, engineResult.usage);
     const sample = evaluateSample({
-        subjectId: workload.subjectId,
+        candidateId: workload.candidateId,
         files: workloadResult.files,
         engineResolveFiles: workload.engineResolveFiles,
         spec: workload.spec,
@@ -1682,7 +1800,7 @@ async function executeAttemptExecutionInCurrentRuntime(args, execution, startedA
             ok: true,
             executionId: execution.id,
             purpose: execution.purpose,
-            subjectId: workload.subjectId,
+            candidateId: workload.candidateId,
             attemptIndex: workload.attemptIndex,
             sampleIndex: workload.sampleIndex,
             caseId: workload.caseId,
@@ -1725,7 +1843,7 @@ export async function executeRuntimeControlOperationSequenceInCurrentRuntime(arg
                 ? { adapterAuthEnv: adapterAuth.env }
                 : {}),
         }, workload, args.runtimeControlOperation.operations.map((operation, index) => runtimeControlStepForOperation(operation, index, args.adapterManifests)), startedAt, {
-            runSubjectPrepare: args.runtimeControlOperation.prepare ?? false,
+            runCandidatePrepare: args.runtimeControlOperation.prepare ?? false,
             workspaceFiles: args.runtimeControlOperation.inputs?.workspace ?? [],
             outputFiles: args.runtimeControlOperation.inputs?.output ?? [],
             collectWorkspace: args.runtimeControlOperation.collectWorkspace ?? false,
@@ -1823,7 +1941,7 @@ function createRuntimeControlSandboxInput(args, request) {
     const parentInput = asRuntimeRecord(args.job.input);
     const publicFiles = runtimeControlInputFiles(request.inputs, "case", parentWorkload.engineCase ? engineCasePublicFiles(parentWorkload.engineCase) : []);
     const privateFiles = runtimeControlInputFiles(request.inputs, "enginePrivate", parentWorkload.engineCase ? engineCasePrivateFiles(parentWorkload.engineCase) : []);
-    const subjectFiles = runtimeControlInputFiles(request.inputs, "subject", parentWorkload.subjectFiles);
+    const candidateFiles = runtimeControlInputFiles(request.inputs, "candidate", parentWorkload.candidateFiles);
     const traceFiles = runtimeControlInputFiles(request.inputs, "traces", parentWorkload.traceFiles);
     const adapter = request.operations[request.operations.length - 1]?.invocation;
     const childExecution = {
@@ -1866,7 +1984,7 @@ function createRuntimeControlSandboxInput(args, request) {
     const childArgs = {
         ...args,
         job: childJob,
-        baseFiles: subjectFiles,
+        baseFiles: candidateFiles,
         engineResolveFiles: [...publicFiles, ...privateFiles],
         engineCases: [engineCase],
         traceFiles,
@@ -1890,10 +2008,10 @@ function runtimeControlStepForOperation(operation, index, manifests = []) {
             ...(operation.invocation.auth !== undefined ? { auth: operation.invocation.auth } : {}),
         }, operation.operation, manifests).command;
     return {
-        kind: operation.operation === "subject.run"
-            ? "subject"
-            : operation.operation === "optimizer.improve"
-                ? "optimizer"
+        kind: operation.operation === "candidate.run"
+            ? "candidate"
+            : operation.operation === "candidate.improve"
+                ? "improver"
                 : "engine",
         label: operation.label ?? `${operation.operation.replace(".", "_")}_${index + 1}`,
         operation: operation.operation,
@@ -1960,8 +2078,8 @@ function isWorkbenchAdapterOperationResult(value) {
     return record.protocol === "workbench.adapter-result.v1" &&
         (record.operation === "engine.resolve" ||
             record.operation === "engine.run" ||
-            record.operation === "subject.run" ||
-            record.operation === "optimizer.improve");
+            record.operation === "candidate.run" ||
+            record.operation === "candidate.improve");
 }
 function cloneSurfaceFiles(files) {
     return files.map((file) => ({ ...file, path: normalizeRelativePath(file.path) }));
@@ -2040,9 +2158,11 @@ async function runHostedCommandExecutionSteps(args, workload, steps, startedAt,
             const stepTimeoutMs = environmentVersion
                 ? environmentVersionTimeoutMs(environmentVersion)
                 : 5 * 60 * 1000;
-            const shouldRunSubjectPrepare = options.runSubjectPrepare ?? steps.some((step) => step.executor === "sandbox");
-            if (shouldRunSubjectPrepare) {
-                await runSubjectPrepareCommand({
+            const shouldRunCandidatePrepare = options.runCandidatePrepare ??
+                (readWorkloadExecutionPurpose(workload) === "attempt" &&
+                    steps.some((step) => step.executor === "sandbox"));
+            if (shouldRunCandidatePrepare) {
+                await runCandidatePrepareCommand({
                     root: workspace.root,
                     workload,
                     execution,
@@ -2081,6 +2201,9 @@ async function runHostedCommandExecutionSteps(args, workload, steps, startedAt,
                     });
                     const operationResult = await readWorkbenchAdapterOperationResult(outputDir(workspace.root), step.operation);
                     assertWorkbenchAdapterOperationResultOk(operationResult, `Adapter ${step.adapter?.use ?? execution.adapter.use} ${step.operation}`);
+                    await writeSurfaceFiles(outputDir(workspace.root), [
+                        textSurfaceFile(`.workbench/traces/${workload.job.id}/${step.label}/result.json`, `${JSON.stringify(operationResult, null, 2)}\n`),
+                    ]);
                     operationResults.push(operationResult);
                     await publishCommandStepEvent(options.eventPublisher, {
                         step: step.label,
@@ -2132,19 +2255,19 @@ async function runHostedCommandExecutionSteps(args, workload, steps, startedAt,
         await workspace.cleanup();
     }
 }
-async function runSubjectPrepareCommand(args) {
-    const command = args.workload.spec.subject.prepare?.command;
+async function runCandidatePrepareCommand(args) {
+    const command = args.workload.spec.candidate.prepare?.command;
     if (!command) {
         return;
     }
-    const role = args.execution.purpose === "improve" ? "optimizer" : "runner";
+    const role = args.execution.purpose === "improve" ? "improver" : "runner";
     await publishCommandStepEvent(args.eventPublisher, {
-        step: "subject_prepare",
+        step: "candidate_prepare",
         status: "started",
         role,
     });
     try {
-        const shellCommand = createHostedWorkloadShellCommand(args.root, command, "subject_prepare");
+        const shellCommand = createHostedWorkloadShellCommand(args.root, command, "candidate_prepare");
         await args.execFileAsync("sh", ["-c", shellCommand], {
             cwd: args.root,
             env: createHostedWorkloadPrepareEnv(args.root),
@@ -2152,20 +2275,20 @@ async function runSubjectPrepareCommand(args) {
             timeout: args.timeoutMs,
         });
         await publishCommandStepEvent(args.eventPublisher, {
-            step: "subject_prepare",
+            step: "candidate_prepare",
             status: "succeeded",
             role,
         });
     }
     catch (error) {
         await publishCommandStepEvent(args.eventPublisher, {
-            step: "subject_prepare",
+            step: "candidate_prepare",
             status: "failed",
             exitCode: readExitCode(error),
             error: error instanceof Error ? error.message : String(error),
             role,
         });
-        throw new Error(`Subject prepare command failed: ${error instanceof Error ? error.message : String(error)}`);
+        throw new Error(`Candidate prepare command failed: ${error instanceof Error ? error.message : String(error)}`);
     }
 }
 async function createRuntimeWorkspaceRoot(args, fs, os, path, prefix) {
@@ -2204,10 +2327,10 @@ async function createRuntimeWorkspaceRoot(args, fs, os, path, prefix) {
     };
 }
 function stepEventRole(step) {
-    if (step.kind === "optimizer") {
-        return "optimizer";
+    if (step.kind === "improver") {
+        return "improver";
     }
-    if (step.kind === "subject") {
+    if (step.kind === "candidate") {
         return "runner";
     }
     if (step.kind === "engine") {
@@ -2219,10 +2342,10 @@ function adapterOperationUsageSummary(result) {
     if (hasExplicitUsageRole(result.usage)) {
         return completeUsageSummary(result.usage);
     }
-    if (result.operation === "optimizer.improve") {
-        return assignUsageRole("optimizer", result.usage);
+    if (result.operation === "candidate.improve") {
+        return assignUsageRole("improver", result.usage);
     }
-    if (result.operation === "subject.run") {
+    if (result.operation === "candidate.run") {
         return assignUsageRole("runner", result.usage);
     }
     if (result.operation === "engine.run") {
@@ -2239,16 +2362,16 @@ function attemptUsageSummary(workloadUsage, resultUsage) {
 }
 function hasExplicitUsageRole(usage) {
     const normalized = completeUsageSummary(usage);
-    return Boolean(normalized?.optimizer || normalized?.runner || normalized?.engine);
+    return Boolean(normalized?.improver || normalized?.runner || normalized?.engine);
 }
-function createSubjectPatchFromResult(result, spec) {
-    if (result.subjectPatch) {
-        return result.subjectPatch;
+function createCandidatePatchFromResult(result, spec) {
+    if (result.candidatePatch) {
+        return result.candidatePatch;
     }
     const changedEditPaths = result.fileChanges
         .map(normalizeRelativePath)
         .filter((filePath) => !filePath.startsWith(".workbench/") &&
-        isSubjectEditPath(filePath, optimizerEdits(spec)));
+        isCandidateEditPath(filePath, improveEdits(spec)));
     const changedSet = new Set(changedEditPaths);
     const files = result.files
         .filter((file) => changedSet.has(normalizeRelativePath(file.path)))
@@ -2260,7 +2383,7 @@ function createSubjectPatchFromResult(result, spec) {
         ...(result.feedback !== undefined ? { feedback: result.feedback } : {}),
     };
 }
-function isSubjectEditPath(filePath, edits) {
+function isCandidateEditPath(filePath, edits) {
     const normalized = normalizeRelativePath(filePath);
     return edits.some((entry) => {
         const editPath = normalizeRelativePath(entry).replace(/\/+$/u, "");
@@ -2320,21 +2443,33 @@ export async function stageWorkbenchRunWorkload(root, workload) {
     ]);
     await fs.mkdir(inputDir(root), { recursive: true });
     await fs.mkdir(outputDir(root), { recursive: true });
+    await clearMutableWorkspaceFiles(root);
     if (purpose === "attempt") {
-        await fs.mkdir(subjectDir(root), { recursive: true });
+        await fs.mkdir(candidateDir(root), { recursive: true });
         await fs.mkdir(caseDir(root), { recursive: true });
         const engineCase = requireWorkloadEngineCase(workload, "Attempt staging");
-        await writeSurfaceFiles(subjectDir(root), workload.subjectFiles);
+        await writeSurfaceFiles(candidateDir(root), workload.candidateFiles);
         await writeSurfaceFiles(caseDir(root), engineCasePublicFiles(engineCase));
         return;
     }
     if (purpose === "improve") {
-        await fs.mkdir(subjectDir(root), { recursive: true });
-        await writeSurfaceFiles(subjectDir(root), workload.subjectFiles);
+        await writeSurfaceFiles(root, workload.candidateFiles.filter((file) => isMutableWorkspaceSnapshotPath(file.path)));
         await fs.mkdir(tracesDir(root), { recursive: true });
         await writeSurfaceFiles(tracesDir(root), workload.traceFiles);
     }
 }
+async function clearMutableWorkspaceFiles(root) {
+    const fs = await importNodeModule(nodeBuiltin("fs/promises"));
+    const path = await importNodeModule(nodeBuiltin("path"));
+    const entries = await fs.readdir(root, { withFileTypes: true }).catch(() => []);
+    await Promise.all(entries.map(async (entry) => {
+        const relativePath = normalizeRelativePath(entry.name);
+        if (!isMutableWorkspaceSnapshotPath(relativePath)) {
+            return;
+        }
+        await fs.rm(path.join(root, entry.name), { recursive: true, force: true });
+    }));
+}
 async function stageWorkbenchEnginePrivateFiles(root, workload) {
     if (readWorkloadExecutionPurpose(workload) !== "attempt") {
         return;
@@ -2417,7 +2552,7 @@ function adapterFilePathWithinRoot(filePath, sourceRoot) {
 }
 async function readHostedRunFailureResult(root, workload, options) {
     const traceFiles = await readRuntimeTraceFiles(root, workload);
-    const outputFiles = filterRuntimeOutputFiles(await readSurfaceFiles(outputDir(root)));
+    const outputFiles = filterRuntimeOutputFiles(await readSurfaceFiles(outputDir(root), { ignorePath: isWorkbenchInternalOutputPath }));
     const startedAt = options.startedAt ?? new Date().toISOString();
     const finishedAt = new Date().toISOString();
     const files = [...outputFiles, ...traceFiles].sort((left, right) => left.path.localeCompare(right.path));
@@ -2433,13 +2568,13 @@ async function readHostedRunFailureResult(root, workload, options) {
 async function readWorkbenchRunWorkloadResult(root, workload, options = {}) {
     const path = await importNodeModule(nodeBuiltin("path"));
     const traceFiles = await readRuntimeTraceFiles(root, workload);
-    const outputFiles = filterRuntimeOutputFiles(await readSurfaceFiles(outputDir(root)));
+    const outputFiles = filterRuntimeOutputFiles(await readSurfaceFiles(outputDir(root), { ignorePath: isWorkbenchInternalOutputPath }));
     const outputExitCode = await readOptionalNumber(path.join(outputDir(root), "exit_code"));
     const startedAt = options.startedAt ?? new Date().toISOString();
     const finishedAt = new Date().toISOString();
     const purpose = readWorkloadExecutionPurpose(workload);
     const primaryOperation = purpose === "improve"
-        ? "optimizer.improve"
+        ? "candidate.improve"
         : "engine.run";
     const primaryResult = [...(options.operationResults ?? [])]
         .reverse()
@@ -2453,9 +2588,9 @@ async function readWorkbenchRunWorkloadResult(root, workload, options = {}) {
     const cases = normalizeResultCases(resultPayload.cases);
     const includeResultScoring = purpose === "attempt";
     const files = [...outputFiles, ...traceFiles].sort((left, right) => left.path.localeCompare(right.path));
-    const subjectPatch = purpose === "improve" ? primaryResult?.value : undefined;
+    const candidatePatch = purpose === "improve" ? primaryResult?.value : undefined;
     const engineResult = purpose === "attempt" ? primaryResult?.value : undefined;
-    const declaredChanges = subjectPatch?.fileChanges ??
+    const declaredChanges = candidatePatch?.fileChanges ??
         (Array.isArray(resultPayload.fileChanges)
             ? resultPayload.fileChanges.filter((entry) => typeof entry === "string")
             : files.map((file) => file.path));
@@ -2463,7 +2598,7 @@ async function readWorkbenchRunWorkloadResult(root, workload, options = {}) {
         files,
         fileChanges: declaredChanges,
         ...(options.operationResults ? { operationResults: [...options.operationResults] } : {}),
-        ...(subjectPatch ? { subjectPatch } : {}),
+        ...(candidatePatch ? { candidatePatch } : {}),
         ...(engineResult ? { result: engineResult } : {}),
         ...(includeResultScoring && metrics ? { metrics } : {}),
         ...(includeResultScoring && cases ? { cases } : {}),
@@ -2537,8 +2672,8 @@ async function writeWorkbenchAdapterRequest(root, workload, execution, step, aut
     await fs.mkdir(path.dirname(requestPath), { recursive: true });
     const casePrompt = workload.engineCaseSpec?.prompt;
     const adapter = step.adapter ?? execution.adapter;
-    const subjectCommand = adapterProtocolCommandSpec(workload.spec.run, "subject.run", manifests).command;
-    await fs.writeFile(requestPath, `${JSON.stringify({
+    const candidateCommand = adapterProtocolCommandSpec(workload.spec.run, "candidate.run", manifests).command;
+    const payload = {
         protocol: "workbench.adapter.v3",
         id: execution.id,
         jobId: workload.job.id,
@@ -2554,17 +2689,17 @@ async function writeWorkbenchAdapterRequest(root, workload, execution, step, aut
                 name: workload.spec.benchmark.name,
                 description: workload.spec.benchmark.description,
             },
-            subject: {
-                id: workload.subjectId,
-                path: workload.spec.subject.files.path,
-                ...(workload.spec.subject.prepare ? { prepare: { ...workload.spec.subject.prepare } } : {}),
+            candidate: {
+                id: workload.candidateId,
+                path: workload.spec.candidate.files.path,
+                ...(workload.spec.candidate.prepare ? { prepare: { ...workload.spec.candidate.prepare } } : {}),
                 run: {
                     ...workload.spec.run,
-                    command: subjectCommand,
+                    command: candidateCommand,
                 },
             },
-            ...(workload.spec.optimizer
-                ? { optimizer: { edits: [...workload.spec.optimizer.edits] } }
+            ...(workload.spec.candidate.improve
+                ? { improve: { edits: [...workload.spec.candidate.improve.edits] } }
                 : {}),
             attempt: {
                 attemptIndex: workload.attemptIndex,
@@ -2580,21 +2715,41 @@ async function writeWorkbenchAdapterRequest(root, workload, execution, step, aut
             workspace: root,
             output: outputDir(root),
             result: workbenchAdapterOperationResultPath(outputDir(root)),
-            subject: subjectDir(root),
+            ...(readWorkloadExecutionPurpose(workload) === "attempt" ? { candidate: candidateDir(root) } : {}),
             ...(workload.engineCaseSpec ? { case: caseDir(root) } : {}),
             traces: tracesDir(root),
             ...(step.kind === "engine" ? { enginePrivate: runtimeEnginePrivateDir(root) } : {}),
         },
-    }, null, 2)}\n`);
+    };
+    await fs.writeFile(requestPath, `${JSON.stringify(payload, null, 2)}\n`);
+    await writeSurfaceFiles(outputDir(root), [
+        textSurfaceFile(`.workbench/traces/${workload.job.id}/${step.label}/request.json`, `${JSON.stringify(sanitizeAdapterRequestTracePayload(payload), null, 2)}\n`),
+    ]);
     return requestPath;
 }
-function optimizerEdits(spec) {
-    return spec.optimizer?.edits ?? [];
+function sanitizeAdapterRequestTracePayload(value) {
+    if (Array.isArray(value)) {
+        return value.map((entry) => sanitizeAdapterRequestTracePayload(entry));
+    }
+    if (!value || typeof value !== "object") {
+        return (value ?? null);
+    }
+    const sanitized = {};
+    for (const [key, entry] of Object.entries(value)) {
+        if (key === "auth" || key === "enginePrivate") {
+            continue;
+        }
+        sanitized[key] = sanitizeAdapterRequestTracePayload(entry);
+    }
+    return sanitized;
+}
+function improveEdits(spec) {
+    return spec.candidate.improve?.edits ?? [];
 }
-function requireOptimizerEdits(spec) {
-    const edits = optimizerEdits(spec);
+function requireImproveEdits(spec) {
+    const edits = improveEdits(spec);
     if (edits.length === 0) {
-        throw new Error("Optimizer YAML must declare at least one entry in edits.");
+        throw new Error("Candidate improve configuration must declare at least one entry in edits.");
     }
     return edits;
 }
@@ -2691,8 +2846,8 @@ function requireWorkloadEngineCase(workload, label) {
     }
     return workload.engineCase;
 }
-function subjectDir(root) {
-    return `${inputDir(root)}/subject`;
+function candidateDir(root) {
+    return `${inputDir(root)}/candidate`;
 }
 function caseDir(root) {
     return `${inputDir(root)}/case`;
@@ -2727,7 +2882,7 @@ async function writeSurfaceFiles(root, files) {
         }
     }
 }
-async function readSurfaceFiles(root) {
+async function readSurfaceFiles(root, options = {}) {
     const fs = await importNodeModule(nodeBuiltin("fs/promises"));
     const path = await importNodeModule(nodeBuiltin("path"));
     const utf8Decoder = new TextDecoder("utf-8", { fatal: true });
@@ -2738,6 +2893,10 @@ async function readSurfaceFiles(root) {
             .catch(() => []);
         for (const entry of entries) {
             const absolutePath = path.join(directory, entry.name);
+            const relativePath = normalizeRelativePath(path.relative(root, absolutePath).replace(/\\/gu, "/"));
+            if (options.ignorePath?.(relativePath)) {
+                continue;
+            }
             if (entry.isDirectory()) {
                 await walk(absolutePath);
                 continue;
@@ -2745,9 +2904,18 @@ async function readSurfaceFiles(root) {
             if (!entry.isFile()) {
                 continue;
             }
-            const relativePath = normalizeRelativePath(path.relative(root, absolutePath).replace(/\\/gu, "/"));
-            const body = await fs.readFile(absolutePath);
-            const stats = await fs.stat(absolutePath);
+            let body;
+            let stats;
+            try {
+                body = await fs.readFile(absolutePath);
+                stats = await fs.stat(absolutePath);
+            }
+            catch (error) {
+                if (isVanishedWalkEntry(error)) {
+                    continue;
+                }
+                throw error;
+            }
             const content = encodeSurfaceSnapshotContent(body, utf8Decoder);
             files.push({
                 path: relativePath,
@@ -2761,6 +2929,10 @@ async function readSurfaceFiles(root) {
     await walk(root);
     return files.sort((left, right) => left.path.localeCompare(right.path));
 }
+function isVanishedWalkEntry(error) {
+    const code = error?.code;
+    return code === "ENOENT" || code === "ENOTDIR";
+}
 function encodeSurfaceSnapshotContent(body, utf8Decoder) {
     try {
         return {
@@ -2943,7 +3115,13 @@ function evaluateSample(args) {
     if (metrics.score === undefined) {
         metrics.score = sampleScore;
     }
-    const cases = args.workload.cases?.length ? args.workload.cases : undefined;
+    const cases = runtimeTimedCaseResults({
+        caseId: args.caseId,
+        status: "completed",
+        durationMs,
+        metrics,
+        cases: args.workload.cases,
+    });
     const feedback = {
         ...(args.workload.summary !== undefined
             ? { summary: args.workload.summary }
@@ -2956,10 +3134,10 @@ function evaluateSample(args) {
     return {
         id: `${args.caseId}__sample_${String(args.sampleIndex + 1).padStart(3, "0")}`,
         index: args.sampleIndex,
-        subject: {
-            id: args.subjectId,
-            kind: "subject",
-            label: args.subjectId,
+        candidate: {
+            id: args.candidateId,
+            kind: "candidate",
+            label: args.candidateId,
         },
         status: "completed",
         startedAt: args.startedAt,
@@ -2967,7 +3145,7 @@ function evaluateSample(args) {
         durationMs,
         metrics,
         ...(usage ? { usage } : {}),
-        ...(cases ? { cases } : {}),
+        cases,
         feedback,
     };
 }
@@ -2976,7 +3154,7 @@ function normalizeSampleJobOutput(value) {
         return null;
     }
     const record = value;
-    if (record.ok !== true || typeof record.subjectId !== "string") {
+    if (record.ok !== true || typeof record.candidateId !== "string") {
         return null;
     }
     const files = Array.isArray(record.files)
@@ -2991,7 +3169,7 @@ function normalizeSampleJobOutput(value) {
         return null;
     }
     return {
-        subjectId: record.subjectId,
+        candidateId: record.candidateId,
         attemptIndex: record.attemptIndex,
         sample,
         fileChanges: Array.isArray(record.fileChanges)
@@ -3006,9 +3184,57 @@ function normalizeSampleJobOutput(value) {
 function normalizeEvaluationSampleOutputs(args) {
     return args.jobs.flatMap((job) => {
         const output = normalizeSampleJobOutput(job.output);
-        return output ? [{ jobs: [job], output }] : [];
+        if (!output) {
+            return [];
+        }
+        const caseId = readJobString(job.input, "caseId") ?? output.sample.cases?.[0]?.id ?? null;
+        const durationMs = runtimeJobDurationMs(job) ?? output.sample.durationMs;
+        const sample = caseId && typeof durationMs === "number" && Number.isFinite(durationMs)
+            ? {
+                ...output.sample,
+                cases: runtimeTimedCaseResults({
+                    caseId,
+                    status: output.sample.status === "error" ? "error" : "completed",
+                    durationMs,
+                    metrics: output.sample.metrics ?? {},
+                    cases: output.sample.cases,
+                }),
+            }
+            : output.sample;
+        return [{
+                jobs: [job],
+                output: {
+                    ...output,
+                    sample,
+                },
+            }];
     });
 }
+function runtimeTimedCaseResults(args) {
+    const cases = args.cases?.length
+        ? args.cases
+        : [{
+                id: args.caseId,
+                status: args.status,
+                metrics: args.metrics,
+            }];
+    return cases.map((entry) => ({
+        ...entry,
+        status: entry.status ?? args.status,
+        metrics: entry.metrics ?? args.metrics,
+        durationMs: args.durationMs,
+    }));
+}
+function runtimeJobDurationMs(job) {
+    if (typeof job.startedAt !== "string" || typeof job.finishedAt !== "string") {
+        return undefined;
+    }
+    const startedMs = Date.parse(job.startedAt);
+    const finishedMs = Date.parse(job.finishedAt);
+    return Number.isFinite(startedMs) && Number.isFinite(finishedMs)
+        ? Math.max(0, finishedMs - startedMs)
+        : undefined;
+}
 function meanFinite(values) {
     const finite = values.filter((value) => typeof value === "number" && Number.isFinite(value));
     if (finite.length === 0) {
@@ -3039,12 +3265,12 @@ function withJobUsage(sample, _jobs, attemptJob) {
         usage,
     };
 }
-function normalizeSubjectRevisionJobOutput(value) {
+function normalizeCandidateRevisionJobOutput(value) {
     if (!value || typeof value !== "object" || Array.isArray(value)) {
         return null;
     }
     const record = value;
-    if (record.ok !== true || typeof record.subjectId !== "string") {
+    if (record.ok !== true || typeof record.candidateId !== "string") {
         return null;
     }
     const files = Array.isArray(record.files)
@@ -3056,7 +3282,7 @@ function normalizeSubjectRevisionJobOutput(value) {
     }
     const usage = normalizeUsageSummary(record.usage);
     return {
-        subjectId: record.subjectId,
+        candidateId: record.candidateId,
         attemptIndex: record.attemptIndex,
         baseId: typeof record.baseId === "string" && record.baseId.length > 0
             ? record.baseId
@@ -3072,7 +3298,7 @@ function normalizeSubjectRevisionJobOutput(value) {
         ...(usage ? { usage } : {}),
     };
 }
-function errorEvaluationSamplesFromJobs(jobs, subjectId, attemptIndex, completedSampleKeys) {
+function errorEvaluationSamplesFromJobs(jobs, candidateId, attemptIndex, completedSampleKeys) {
     const groups = new Map();
     for (const job of jobs) {
         const key = evaluationSampleGroupKeyFromJob(job);
@@ -3082,10 +3308,10 @@ function errorEvaluationSamplesFromJobs(jobs, subjectId, attemptIndex, completed
         groups.set(key, [...(groups.get(key) ?? []), job]);
     }
     return [...groups.values()]
-        .map((group) => errorEvaluationSampleFromJobGroup(group, subjectId, attemptIndex))
+        .map((group) => errorEvaluationSampleFromJobGroup(group, candidateId, attemptIndex))
         .filter((sample) => sample !== null);
 }
-function errorEvaluationSampleFromJobGroup(jobs, subjectId, attemptIndex) {
+function errorEvaluationSampleFromJobGroup(jobs, candidateId, attemptIndex) {
     const job = jobs[0];
     if (!job) {
         return null;
@@ -3097,25 +3323,27 @@ function errorEvaluationSampleFromJobGroup(jobs, subjectId, attemptIndex) {
     }
     const startedAt = minIsoTimestamp(jobs.map((entry) => entry.startedAt ?? entry.createdAt));
     const finishedAt = maxIsoTimestamp(jobs.map((entry) => entry.finishedAt ?? entry.updatedAt ?? entry.startedAt));
+    const durationMs = startedAt && finishedAt
+        ? Math.max(0, Date.parse(finishedAt) - Date.parse(startedAt))
+        : undefined;
     const error = summarizeEvaluationJobErrors(jobs) ?? "Evaluation job did not produce a valid sample.";
     return {
         id: `${caseId}__sample_${String(sampleIndex + 1).padStart(3, "0")}`,
         index: sampleIndex,
-        subject: {
-            id: subjectId,
-            kind: "subject",
-            label: subjectId,
+        candidate: {
+            id: candidateId,
+            kind: "candidate",
+            label: candidateId,
         },
         status: "error",
         ...(startedAt ? { startedAt } : {}),
         ...(finishedAt ? { finishedAt } : {}),
-        ...(startedAt && finishedAt
-            ? { durationMs: Math.max(0, Date.parse(finishedAt) - Date.parse(startedAt)) }
-            : {}),
+        ...(durationMs !== undefined ? { durationMs } : {}),
         ...(error ? { error } : {}),
         cases: [{
                 id: caseId,
                 status: "error",
+                ...(durationMs !== undefined ? { durationMs } : {}),
                 metrics: {},
                 ...(error ? { feedback: { summary: error } } : {}),
             }],
@@ -3171,13 +3399,13 @@ function compareSampleOutputs(left, right) {
     }
     return left.sample.id.localeCompare(right.sample.id);
 }
-function createEvaluationRecord(subjectId, subjectName, rawSamples) {
-    const samples = mergeEvaluationSampleRecords(rawSamples).map((sample) => subjectName
+function createEvaluationRecord(candidateId, candidateName, rawSamples) {
+    const samples = mergeEvaluationSampleRecords(rawSamples).map((sample) => candidateName
         ? {
             ...sample,
-            subject: {
-                ...sample.subject,
-                label: subjectName,
+            candidate: {
+                ...sample.candidate,
+                label: candidateName,
             },
         }
         : sample);
@@ -3191,10 +3419,10 @@ function createEvaluationRecord(subjectId, subjectName, rawSamples) {
     const errorSampleCount = samples.filter((sample) => sample.status === "error")
         .length;
     return {
-        subject: {
-            id: subjectId,
-            kind: "subject",
-            ...(subjectName ? { label: subjectName } : {}),
+        candidate: {
+            id: candidateId,
+            kind: "candidate",
+            ...(candidateName ? { label: candidateName } : {}),
         },
         status: samples.length > 0 && completedSampleCount === samples.length
             ? "completed"
@@ -3215,7 +3443,7 @@ function createEvaluationRecord(subjectId, subjectName, rawSamples) {
         samples,
     };
 }
-function normalizedSubjectDisplayName(value) {
+function normalizedCandidateDisplayName(value) {
     const normalized = value?.trim();
     return normalized ? normalized : null;
 }
@@ -3263,7 +3491,7 @@ function mergeEvaluationSampleGroup(group) {
     return {
         id: `sample_${String(first.index + 1).padStart(3, "0")}`,
         index: first.index,
-        subject: first.subject,
+        candidate: first.candidate,
         status: mergeEvaluationSampleStatus(group),
         ...(startedAt ? { startedAt } : {}),
         ...(finishedAt ? { finishedAt } : {}),
@@ -3355,34 +3583,28 @@ function aggregateCaseStatus(results) {
     }
     return undefined;
 }
-function evaluationMeanMetrics(evaluation) {
-    const entries = Object.entries(evaluation.metrics ?? {}).filter((entry) => Number.isFinite(entry[1].mean));
-    return entries.length > 0
-        ? Object.fromEntries(entries.map(([key, stats]) => [key, Number(stats.mean.toFixed(3))]))
-        : undefined;
-}
-function selectSubject(args) {
-    let selected = args.previousSubject;
-    for (const subject of args.subjects) {
-        if (!selected || hasHigherScore(subject, selected)) {
-            selected = subject;
+function selectCandidate(args) {
+    let selected = args.previousCandidate;
+    for (const candidate of args.candidates) {
+        if (!selected || hasHigherScore(candidate, selected)) {
+            selected = candidate;
         }
     }
     return selected;
 }
-function hasHigherScore(subject, incumbent) {
-    const subjectValue = readMetric(subject, "score");
-    const incumbentValue = readMetric(incumbent, "score");
-    if (subjectValue == null) {
+function hasHigherScore(candidate, incumbent) {
+    const candidateValue = readEvaluationMean(candidate.eval, "score");
+    const incumbentValue = readEvaluationMean(incumbent.eval, "score");
+    if (candidateValue == null) {
         return false;
     }
     if (incumbentValue == null) {
         return true;
     }
-    return subjectValue > incumbentValue;
+    return candidateValue > incumbentValue;
 }
-function readMetric(subject, metric) {
-    const direct = subject.metrics?.[metric];
+function readEvaluationMean(evaluation, metric) {
+    const direct = evaluation?.metrics?.[metric]?.mean;
     return typeof direct === "number" && Number.isFinite(direct) ? direct : null;
 }
 function metricStats(values) {
@@ -3501,7 +3723,7 @@ function isEvaluationSampleRecord(value) {
         !Array.isArray(value) &&
         typeof record.id === "string" &&
         typeof record.index === "number" &&
-        typeof record.subject === "object" &&
+        typeof record.candidate === "object" &&
         isEvaluationSampleStatus(record.status) &&
         hasOperationalCaseStatuses(record.cases));
 }