@workbench-ai/workbench-built-in-adapters 0.0.49 → 0.0.51

package/dist/agent-turn.d.ts

@@ -7,7 +7,7 @@ export interface AgentProviderSpec {
     effort?: string;
 }
 export interface WorkbenchAgentTurnRequest {
-    role: "optimizer" | "runner" | "engine";
+    role: "improver" | "runner" | "engine";
     provider: AgentProviderSpec;
     adapterAuthRoot?: string;
     adapterAuthRequest?: JsonValue;
@@ -29,4 +29,11 @@ export interface WorkbenchAgentTurnResult {
 export type WorkbenchAgentTurnExecutor = (request: WorkbenchAgentTurnRequest) => Promise<WorkbenchAgentTurnResult>;
 export declare function executeWorkbenchAgentTurn(executor: (request: WorkbenchAgentTurnRequest) => Promise<WorkbenchAgentTurnResult>, request: WorkbenchAgentTurnRequest): Promise<WorkbenchAgentTurnResult>;
 export declare function defaultWorkbenchAgentTurnExecutor(request: WorkbenchAgentTurnRequest): Promise<WorkbenchAgentTurnResult>;
+export declare function resolveAgentTurnTimeouts(defaults: {
+    turn_timeout_ms?: number;
+    stall_timeout_ms?: number;
+}): {
+    turnTimeoutMs: number;
+    stallTimeoutMs: number;
+};
 //# sourceMappingURL=agent-turn.d.ts.map

package/dist/agent-turn.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-turn.d.ts","sourceRoot":"","sources":["../src/agent-turn.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAEV,mBAAmB,EACnB,YAAY,EACb,MAAM,kCAAkC,CAAC;AAC1C,OAAO,EAQL,KAAK,SAAS,EAEf,MAAM,4BAA4B,CAAC;AACpC,OAAO,KAAK,EACV,gCAAgC,EACjC,MAAM,8BAA8B,CAAC;AAetC,MAAM,WAAW,iBAAiB;IAChC,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,yBAAyB;IACxC,IAAI,EAAE,WAAW,GAAG,QAAQ,GAAG,QAAQ,CAAC;IACxC,QAAQ,EAAE,iBAAiB,CAAC;IAC5B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,SAAS,CAAC;IAC/B,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACxC,aAAa,EAAE,MAAM,CAAC;IACtB,GAAG,EAAE,MAAM,CAAC;IACZ,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,cAAc,CAAC,EAAE,gCAAgC,CAAC;CACnD;AAED,MAAM,WAAW,wBAAwB;IACvC,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,mBAAmB,EAAE,CAAC;IAClC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;IACpC,KAAK,CAAC,EAAE,YAAY,CAAC;CACtB;AAED,MAAM,MAAM,0BAA0B,GAAG,CAAC,OAAO,EAAE,yBAAyB,KAAK,OAAO,CAAC,wBAAwB,CAAC,CAAC;AA4BnH,wBAAsB,yBAAyB,CAC7C,QAAQ,EAAE,CAAC,OAAO,EAAE,yBAAyB,KAAK,OAAO,CAAC,wBAAwB,CAAC,EACnF,OAAO,EAAE,yBAAyB,GACjC,OAAO,CAAC,wBAAwB,CAAC,CAenC;AAED,wBAAsB,iCAAiC,CACrD,OAAO,EAAE,yBAAyB,GACjC,OAAO,CAAC,wBAAwB,CAAC,CAwFnC"}
+{"version":3,"file":"agent-turn.d.ts","sourceRoot":"","sources":["../src/agent-turn.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAEV,mBAAmB,EACnB,YAAY,EACb,MAAM,kCAAkC,CAAC;AAC1C,OAAO,EAQL,KAAK,SAAS,EAEf,MAAM,4BAA4B,CAAC;AACpC,OAAO,KAAK,EACV,gCAAgC,EACjC,MAAM,8BAA8B,CAAC;AAiBtC,MAAM,WAAW,iBAAiB;IAChC,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,yBAAyB;IACxC,IAAI,EAAE,UAAU,GAAG,QAAQ,GAAG,QAAQ,CAAC;IACvC,QAAQ,EAAE,iBAAiB,CAAC;IAC5B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,SAAS,CAAC;IAC/B,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACxC,aAAa,EAAE,MAAM,CAAC;IACtB,GAAG,EAAE,MAAM,CAAC;IACZ,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,cAAc,CAAC,EAAE,gCAAgC,CAAC;CACnD;AAED,MAAM,WAAW,wBAAwB;IACvC,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,mBAAmB,EAAE,CAAC;IAClC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;IACpC,KAAK,CAAC,EAAE,YAAY,CAAC;CACtB;AAED,MAAM,MAAM,0BAA0B,GAAG,CAAC,OAAO,EAAE,yBAAyB,KAAK,OAAO,CAAC,wBAAwB,CAAC,CAAC;AA4BnH,wBAAsB,yBAAyB,CAC7C,QAAQ,EAAE,CAAC,OAAO,EAAE,yBAAyB,KAAK,OAAO,CAAC,wBAAwB,CAAC,EACnF,OAAO,EAAE,yBAAyB,GACjC,OAAO,CAAC,wBAAwB,CAAC,CAenC;AAED,wBAAsB,iCAAiC,CACrD,OAAO,EAAE,yBAAyB,GACjC,OAAO,CAAC,wBAAwB,CAAC,CAwFnC;AAkHD,wBAAgB,wBAAwB,CAAC,QAAQ,EAAE;IACjD,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B,GAAG;IACF,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;CACxB,CAQA"}

package/dist/agent-turn.js

@@ -8,6 +8,8 @@ import { importWorkbenchRuntime } from "./runtime.js";
 const DEFAULT_AGENT_TURN_MAX_ATTEMPTS = 3;
 const DEFAULT_AGENT_TURN_RETRY_BASE_MS = 5_000;
 const DEFAULT_AGENT_TURN_RETRY_MAX_MS = 30_000;
+const DEFAULT_AGENT_TURN_TIMEOUT_MS = 3_600_000;
+const DEFAULT_AGENT_STALL_TIMEOUT_MS = 300_000;
 const AGENT_PROVIDER_REGISTRY = {
     codex: {
         executable: "codex",
@@ -201,14 +203,14 @@ function agentProviderRegistration(providerName) {
     return registration;
 }
 async function buildAgentExecutionPlan(provider, providerSpec, workspaceRoot, agentHome, adapterAuth) {
-    const turnTimeoutMs = provider.manifest.defaults.turn_timeout_ms ?? 3_600_000;
+    const { turnTimeoutMs, stallTimeoutMs } = resolveAgentTurnTimeouts(provider.manifest.defaults);
     const harness = {
         id: provider.manifest.id,
         auth: await resolveAgentAuth(provider, providerSpec, workspaceRoot, agentHome, adapterAuth),
         ...(firstNonEmpty(providerSpec.model, provider.manifest.defaults.model) ? { model: firstNonEmpty(providerSpec.model, provider.manifest.defaults.model) } : {}),
         ...(firstNonEmpty(providerSpec.effort, provider.manifest.defaults.effort) ? { effort: firstNonEmpty(providerSpec.effort, provider.manifest.defaults.effort) } : {}),
         turn_timeout_ms: turnTimeoutMs,
-        stall_timeout_ms: Math.max(provider.manifest.defaults.stall_timeout_ms ?? 0, turnTimeoutMs),
+        stall_timeout_ms: stallTimeoutMs,
         config: resolveAgentConfig(provider, defaultWorkbenchAgentConfig(provider, providerSpec.use)),
         retry: DEFAULT_HARNESS_RETRY,
         cancel: DEFAULT_HARNESS_CANCEL,
@@ -221,6 +223,19 @@ async function buildAgentExecutionPlan(provider, providerSpec, workspaceRoot, ag
         harness,
     };
 }
+export function resolveAgentTurnTimeouts(defaults) {
+    const turnTimeoutMs = positiveTimeoutMs(defaults.turn_timeout_ms) ?? DEFAULT_AGENT_TURN_TIMEOUT_MS;
+    const requestedStallTimeoutMs = positiveTimeoutMs(defaults.stall_timeout_ms) ?? DEFAULT_AGENT_STALL_TIMEOUT_MS;
+    return {
+        turnTimeoutMs,
+        stallTimeoutMs: Math.min(requestedStallTimeoutMs, turnTimeoutMs),
+    };
+}
+function positiveTimeoutMs(value) {
+    return typeof value === "number" && Number.isFinite(value) && value > 0
+        ? value
+        : null;
+}
 function defaultWorkbenchAgentConfig(provider, providerName) {
     const fallback = (provider.manifest.defaults.config ?? {});
     return {
@@ -229,9 +244,9 @@ function defaultWorkbenchAgentConfig(provider, providerName) {
     };
 }
 async function resolveAgentAuth(provider, providerSpec, workspaceRoot, agentHome, adapterAuth) {
-    const subject = adapterAuthProviderSubject(adapterAuth.request, providerSpec.use) ??
+    const candidate = adapterAuthProviderCandidate(adapterAuth.request, providerSpec.use) ??
         (provider.manifest.defaults.auth ?? {});
-    const parsed = provider.schemas.auth.safeParse(subject);
+    const parsed = provider.schemas.auth.safeParse(candidate);
     if (!parsed.success) {
         throw new Error(`Agent provider "${provider.manifest.id}" auth is invalid: ${formatValidationIssues(parsed.error.issues)}`);
     }
@@ -239,7 +254,7 @@ async function resolveAgentAuth(provider, providerSpec, workspaceRoot, agentHome
     void agentHome;
     return { ...parsed.data };
 }
-function adapterAuthProviderSubject(auth, providerName) {
+function adapterAuthProviderCandidate(auth, providerName) {
     const record = jsonRecord(auth);
     const self = jsonRecord(record?.self);
     const adapters = jsonRecord(record?.adapters);
@@ -338,7 +353,7 @@ function isTransientAgentTurnError(error) {
     if (isNativeCaCertificateFailure(message)) {
         return false;
     }
-    return /\b(fetch failed|error sending request|stream disconnected before completion|ECONNRESET|ETIMEDOUT|EAI_AGAIN|ENETUNREACH|ECONNREFUSED|socket hang up|network error|UND_ERR_|signal SIGTERM)/iu.test(message);
+    return /\b(fetch failed|error sending request|stream disconnected before completion|turn stalled after \d+ms|ECONNRESET|ETIMEDOUT|EAI_AGAIN|ENETUNREACH|ECONNREFUSED|socket hang up|network error|UND_ERR_|signal SIGTERM)/iu.test(message);
 }
 function isNativeCaCertificateFailure(message) {
     return /\bno native root CA certificates found\b|install ca-certificates/iu.test(message);

package/dist/execute.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"execute.d.ts","sourceRoot":"","sources":["../src/execute.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAEV,IAAI,EAKL,MAAM,kCAAkC,CAAC;AAc1C,OAAO,KAAK,EAEV,0BAA0B,EAG3B,MAAM,iBAAiB,CAAC;AAQzB,MAAM,WAAW,4CAA4C;IAC3D,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,aAAa,CAAC,EAAE,0BAA0B,CAAC;IAC3C,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,IAAI,CAAC;IAC1B,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACzC;AA4CD,wBAAsB,qCAAqC,CACzD,IAAI,GAAE,4CAAiD,GACtD,OAAO,CAAC,IAAI,CAAC,CAiEf"}
+{"version":3,"file":"execute.d.ts","sourceRoot":"","sources":["../src/execute.ts"],"names":[],"mappings":"AAKA,OAAO,KAAK,EAEV,IAAI,EAKL,MAAM,kCAAkC,CAAC;AAc1C,OAAO,KAAK,EAEV,0BAA0B,EAG3B,MAAM,iBAAiB,CAAC;AAQzB,MAAM,WAAW,4CAA4C;IAC3D,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,aAAa,CAAC,EAAE,0BAA0B,CAAC;IAC3C,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,IAAI,CAAC;IAC1B,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACzC;AA4CD,wBAAsB,qCAAqC,CACzD,IAAI,GAAE,4CAAiD,GACtD,OAAO,CAAC,IAAI,CAAC,CAiEf"}

package/dist/execute.js

@@ -1,5 +1,6 @@
 import { spawn } from "node:child_process";
 import { promises as fs } from "node:fs";
+import os from "node:os";
 import path from "node:path";
 import { ensureWorkbenchAdapterOutputDir, readWorkbenchAdapterOperationResult, readWorkbenchAdapterOperationRequest, runWorkbenchRuntimeOperationSequence, writeWorkbenchAdapterOperationResult, workbenchAdapterOperationResultPath, } from "@workbench-ai/workbench-protocol";
 import YAML from "yaml";
@@ -47,8 +48,8 @@ export async function executeWorkbenchBuiltInAdapterCommand(args = {}) {
     if (isBuiltInAgentAdapterId(adapterId)) {
         const workload = workloadFromAdapterOperationRequest(request);
         const agent = builtInAgentSpecFromRequest(request);
-        if (request.operation === "optimizer.improve") {
-            await writeAgentSubjectRevisionOutput(request, workload, agent, {
+        if (request.operation === "candidate.improve") {
+            await writeAgentCandidateRevisionOutput(request, workload, agent, {
                 agentExecutor: args.agentExecutor,
                 adapterAuthRoot: args.adapterAuthRoot,
                 adapterAuthRequest: args.adapterAuthRequest ?? request.auth,
@@ -56,8 +57,8 @@ export async function executeWorkbenchBuiltInAdapterCommand(args = {}) {
             });
             return;
         }
-        if (request.operation === "subject.run") {
-            await writeAgentSubjectOutput(request, workload, agent, {
+        if (request.operation === "candidate.run") {
+            await writeAgentCandidateOutput(request, workload, agent, {
                 agentExecutor: args.agentExecutor,
                 adapterAuthRoot: args.adapterAuthRoot,
                 adapterAuthRequest: args.adapterAuthRequest ?? request.auth,
@@ -123,7 +124,7 @@ async function workbenchEngineOutcomeUsage(outcome) {
     const operationUsage = outcome.usage
         ? undefined
         : runtime.mergeUsageSummaries(outcome.operationResults.map((result) => {
-            if (result.operation === "subject.run") {
+            if (result.operation === "candidate.run") {
                 return runtime.assignUsageRole("runner", result.usage);
             }
             if (result.operation === "engine.run") {
@@ -163,16 +164,16 @@ function workbenchEngineScoreInvocation(request) {
             : adapterCommandName(score.use),
     };
 }
-function workbenchEngineSubjectInvocation(request) {
-    const subject = request.context?.subject?.run;
-    if (!subject?.use || !subject.command) {
-        throw new Error("Workbench engine requires context.subject.run.use and context.subject.run.command.");
+function workbenchEngineCandidateInvocation(request) {
+    const candidate = request.context?.candidate?.run;
+    if (!candidate?.use || !candidate.command) {
+        throw new Error("Workbench engine requires context.candidate.run.use and context.candidate.run.command.");
     }
     return {
-        use: subject.use,
-        with: (subject.with ?? {}),
-        ...(subject.auth !== undefined ? { auth: subject.auth } : {}),
-        command: subject.command,
+        use: candidate.use,
+        with: (candidate.with ?? {}),
+        ...(candidate.auth !== undefined ? { auth: candidate.auth } : {}),
+        command: candidate.command,
     };
 }
 function workbenchEngineGradingIsolation(request) {
@@ -188,13 +189,13 @@ function workbenchEngineGradingIsolation(request) {
 }
 async function runWorkbenchEngineSharedGrading(request) {
     const inputs = await workbenchEngineRuntimeInputs(request);
-    const subject = workbenchEngineSubjectInvocation(request);
+    const candidate = workbenchEngineCandidateInvocation(request);
     const score = workbenchEngineScoreInvocation(request);
     const result = await runWorkbenchRuntimeOperationSequence({
         inputs,
         prepare: true,
         operations: [
-            { label: "subject", operation: "subject.run", invocation: subject },
+            { label: "candidate", operation: "candidate.run", invocation: candidate },
             { label: "score", operation: "engine.run", invocation: score },
         ],
     });
@@ -203,25 +204,25 @@ async function runWorkbenchEngineSharedGrading(request) {
 }
 async function runWorkbenchEngineSeparateGrading(request) {
     const inputs = await workbenchEngineRuntimeInputs(request);
-    const subject = workbenchEngineSubjectInvocation(request);
+    const candidate = workbenchEngineCandidateInvocation(request);
     const score = workbenchEngineScoreInvocation(request);
     const runtime = await importWorkbenchRuntime();
     const runner = await runWorkbenchRuntimeOperationSequence({
         inputs: {
-            subject: inputs.subject,
+            candidate: inputs.candidate,
             case: inputs.case,
             traces: inputs.traces,
         },
         prepare: true,
         collectWorkspace: true,
         operations: [
-            { label: "subject", operation: "subject.run", invocation: subject },
+            { label: "candidate", operation: "candidate.run", invocation: candidate },
         ],
     });
     assertRuntimeControlResultOk(runner, "Workbench separate runner");
     const grader = await runWorkbenchRuntimeOperationSequence({
         inputs: {
-            subject: inputs.subject,
+            candidate: inputs.candidate,
             case: inputs.case,
             enginePrivate: inputs.enginePrivate,
             traces: inputs.traces,
@@ -243,14 +244,14 @@ async function runWorkbenchEngineSeparateGrading(request) {
     };
 }
 async function workbenchEngineRuntimeInputs(request) {
-    const [subject, caseFiles, enginePrivate, traces] = await Promise.all([
-        readOptionalSurfaceFiles(request.paths.subject),
+    const [candidate, caseFiles, enginePrivate, traces] = await Promise.all([
+        readOptionalSurfaceFiles(request.paths.candidate),
         readOptionalSurfaceFiles(request.paths.case),
         readOptionalSurfaceFiles(request.paths.enginePrivate),
         readOptionalSurfaceFiles(request.paths.traces),
     ]);
     return {
-        subject,
+        candidate,
         case: caseFiles,
         enginePrivate,
         traces,
@@ -309,12 +310,20 @@ function safeInternalPathSegment(value) {
 }
 async function executeCommandAdapterRequest(request) {
     const command = requiredAdapterCommandString(request, "command");
-    await runAdapterShellCommand(command, request.paths.workspace);
-    if (request.operation === "engine.run") {
-        await requireCommandScoreResult(request);
-        return;
+    const before = request.operation === "candidate.improve"
+        ? await snapshotEditableCandidateWorkspace(request)
+        : null;
+    try {
+        await runAdapterShellCommand(command, request.paths.workspace);
+        if (request.operation === "engine.run") {
+            await requireCommandScoreResult(request);
+            return;
+        }
+        await writeOperationOkUnlessPresent(request, before?.root);
+    }
+    finally {
+        await before?.cleanup();
     }
-    await writeOperationOkUnlessPresent(request);
 }
 async function requireCommandScoreResult(request) {
     if (!await fileExists(workbenchAdapterOperationResultPath(request.paths.output))) {
@@ -379,15 +388,15 @@ async function runAdapterShellCommand(command, cwd, env = {}) {
         });
     });
 }
-async function writeOperationOkUnlessPresent(request) {
+async function writeOperationOkUnlessPresent(request, beforeRoot) {
     if (await fileExists(workbenchAdapterOperationResultPath(request.paths.output))) {
         return;
     }
-    if (request.operation === "optimizer.improve") {
-        const patch = await createSubjectPatchFromWorkspace({
-            beforeRoot: requiredRequestPath(request.paths.subject, "paths.subject"),
+    if (request.operation === "candidate.improve") {
+        const patch = await createCandidatePatchFromWorkspace({
+            beforeRoot: beforeRoot ?? requiredRequestPath(request.paths.candidate, "paths.candidate"),
             afterRoot: request.paths.workspace,
-            edits: request.context?.optimizer?.edits ?? [],
+            edits: request.context?.improve?.edits ?? [],
         });
         await writeWorkbenchAdapterOperationResult(request.paths.output, {
             protocol: "workbench.adapter-result.v1",
@@ -403,6 +412,41 @@ async function writeOperationOkUnlessPresent(request) {
         ok: true,
     });
 }
+async function snapshotEditableCandidateWorkspace(request) {
+    const root = await fs.mkdtemp(path.join(os.tmpdir(), "workbench-candidate-before-"));
+    const edits = request.context?.improve?.edits ?? [];
+    const files = await readEditableCandidateWorkspaceFiles(request.paths.workspace, edits);
+    await writeSurfaceFiles(root, files);
+    return {
+        root,
+        cleanup: async () => {
+            await fs.rm(root, { recursive: true, force: true }).catch(() => undefined);
+        },
+    };
+}
+async function readEditableCandidateWorkspaceFiles(root, edits) {
+    const files = [];
+    for (const edit of edits) {
+        const normalized = normalizeRelativePath(edit);
+        if (!normalized || isRuntimeWorkspacePath(normalized)) {
+            continue;
+        }
+        const absolutePath = path.join(root, normalized);
+        const stat = await fs.stat(absolutePath).catch(() => null);
+        if (!stat) {
+            continue;
+        }
+        if (stat.isDirectory()) {
+            await readSurfaceFilesInto(root, normalized, files);
+            continue;
+        }
+        if (stat.isFile()) {
+            files.push(await readSurfaceFile(root, normalized));
+        }
+    }
+    return dedupeSurfaceFiles(files.filter((file) => isCandidateEditPath(file.path, edits) &&
+        !isRuntimeWorkspacePath(file.path)));
+}
 async function firstExistingFile(files) {
     for (const file of files) {
         const stat = await fs.stat(file).catch(() => null);
@@ -460,10 +504,13 @@ async function readWorkbenchEngineCase(args) {
         throw new Error(`Task ${args.id} ${TASK_CONTROL_FILE} must include a task string.`);
     }
     const unsupportedTaskFields = Object.keys(taskRecord)
-        .filter((key) => !["version", "task", "files", "tests", "solution", "environment"].includes(key));
+        .filter((key) => !["version", "task", "split", "files", "tests", "solution", "environment"].includes(key));
     if (unsupportedTaskFields.length > 0) {
         throw new Error(`Task ${args.id} ${TASK_CONTROL_FILE} has unsupported field${unsupportedTaskFields.length === 1 ? "" : "s"}: ${unsupportedTaskFields.join(", ")}.`);
     }
+    if (taskRecord.split !== undefined && (typeof taskRecord.split !== "string" || taskRecord.split.trim().length === 0)) {
+        throw new Error(`Task ${args.id} ${TASK_CONTROL_FILE} split must be a non-empty string when provided.`);
+    }
     const publicPrefix = taskDirectoryPrefix(taskRecord.files, "files", args.id);
     const testsPrefix = taskDirectoryPrefix(taskRecord.tests, "tests", args.id);
     const solutionPrefix = taskDirectoryPrefix(taskRecord.solution, "solution", args.id);
@@ -483,6 +530,7 @@ async function readWorkbenchEngineCase(args) {
         case: {
             version: 3,
             prompt: taskRecord.task,
+            ...(typeof taskRecord.split === "string" ? { split: taskRecord.split.trim() } : {}),
             ...(taskRecord.environment !== undefined
                 ? { environment: taskRecord.environment }
                 : {}),
@@ -539,21 +587,25 @@ async function readSurfaceFilesInto(root, relativeDir, result) {
         if (!entry.isFile()) {
             continue;
         }
-        const [body, stat] = await Promise.all([
-            fs.readFile(absolutePath),
-            fs.stat(absolutePath),
-        ]);
-        const text = body.toString("utf8");
-        const isUtf8 = Buffer.from(text, "utf8").equals(body);
-        result.push({
-            path: relativePath,
-            kind: isUtf8 ? "text" : "binary",
-            encoding: isUtf8 ? "utf8" : "base64",
-            content: isUtf8 ? text : body.toString("base64"),
-            executable: (stat.mode & 0o111) !== 0,
-        });
+        result.push(await readSurfaceFile(root, relativePath));
     }
 }
+async function readSurfaceFile(root, relativePath) {
+    const absolutePath = path.join(root, normalizeRelativePath(relativePath));
+    const [body, stat] = await Promise.all([
+        fs.readFile(absolutePath),
+        fs.stat(absolutePath),
+    ]);
+    const text = body.toString("utf8");
+    const isUtf8 = Buffer.from(text, "utf8").equals(body);
+    return {
+        path: normalizeRelativePath(relativePath),
+        kind: isUtf8 ? "text" : "binary",
+        encoding: isUtf8 ? "utf8" : "base64",
+        content: isUtf8 ? text : body.toString("base64"),
+        executable: (stat.mode & 0o111) !== 0,
+    };
+}
 async function fileExists(filePath) {
     return fs.stat(filePath).then((stat) => stat.isFile(), () => false);
 }
@@ -644,13 +696,13 @@ function workloadFromAdapterOperationRequest(request) {
             name: context.benchmark?.name ?? "",
             description: context.benchmark?.description ?? "",
         },
-        subject: {
-            path: context.subject?.path ?? "",
+        candidate: {
+            path: context.candidate?.path ?? "",
         },
-        optimizer: {
-            edits: context.optimizer?.edits ?? [],
+        improve: {
+            edits: context.improve?.edits ?? [],
         },
-        subjectId: context.subject?.id ?? "",
+        candidateId: context.candidate?.id ?? "",
         attemptIndex: attempt.attemptIndex ?? 0,
         sampleIndex: attempt.sampleIndex ?? 0,
         caseId: attempt.caseId ?? "",
@@ -726,35 +778,35 @@ async function executeBuiltInAgentTurn(executor, request) {
     const { defaultWorkbenchAgentTurnExecutor, executeWorkbenchAgentTurn, } = await import("./agent-turn.js");
     return await executeWorkbenchAgentTurn(executor ?? defaultWorkbenchAgentTurnExecutor, request);
 }
-async function writeAgentSubjectOutput(request, workload, subject, options = {}) {
-    if (request.operation !== "subject.run") {
-        throw new Error("Agent subject results can only complete subject.run operations.");
+async function writeAgentCandidateOutput(request, workload, candidate, options = {}) {
+    if (request.operation !== "candidate.run") {
+        throw new Error("Agent candidate results can only complete candidate.run operations.");
     }
-    const traceRoot = path.join(request.paths.output, ".workbench", "internal", "agent-subject");
+    const traceRoot = path.join(request.paths.output, ".workbench", "internal", "agent-candidate");
     const agentResult = await executeBuiltInAgentTurn(options.agentExecutor, {
         role: "runner",
-        provider: subject.agent,
+        provider: candidate.agent,
         adapterAuthRoot: options.adapterAuthRoot,
         adapterAuthRequest: options.adapterAuthRequest,
         adapterAuthEnv: options.adapterAuthEnv,
         workspaceRoot: request.paths.workspace,
         cwd: request.paths.workspace,
-        prompt: buildAgentSubjectPrompt(workload, subject),
+        prompt: buildAgentCandidatePrompt(workload, candidate),
         traceRoot,
         jobId: workload.job.id,
     });
-    const outputPath = path.join(request.paths.output, "subject-summary.md");
+    const outputPath = path.join(request.paths.output, "candidate-summary.md");
     await fs.mkdir(path.dirname(outputPath), { recursive: true });
     await fs.writeFile(outputPath, agentResult.output);
     const trace = {
-        path: `.workbench/traces/${workload.job.id}/subject.json`,
+        path: `.workbench/traces/${workload.job.id}/candidate.json`,
         kind: "text",
         encoding: "utf8",
         executable: false,
         content: `${JSON.stringify({
-            kind: "agent_subject",
-            provider: subject.agent.use,
-            subjectId: workload.subjectId,
+            kind: "agent_candidate",
+            provider: candidate.agent.use,
+            candidateId: workload.candidateId,
             attemptIndex: workload.attemptIndex,
             sampleIndex: workload.sampleIndex,
             summary: agentResult.output,
@@ -766,25 +818,25 @@ async function writeAgentSubjectOutput(request, workload, subject, options = {})
     const usage = runtime.assignUsageRole("runner", agentResult.usage);
     await writeWorkbenchAdapterOperationResult(request.paths.output, {
         protocol: "workbench.adapter-result.v1",
-        operation: "subject.run",
+        operation: "candidate.run",
         ok: true,
         ...(agentResult.output ? { summary: agentResult.output } : {}),
         feedback: {
-            subject: "agent",
-            agent: subject.agent.use,
+            candidate: "agent",
+            agent: candidate.agent.use,
             metadata: agentResult.metadata,
         },
         ...(usage ? { usage } : {}),
     });
 }
-function buildAgentSubjectPrompt(workload, subject) {
+function buildAgentCandidatePrompt(workload, candidate) {
     return [
-        ...(subject.instructions ? ["Instructions:", subject.instructions, ""] : []),
+        ...(candidate.instructions ? ["Instructions:", candidate.instructions, ""] : []),
         "Context:",
-        "- Subject source files are mounted at /workspace/input/subject.",
-        "- Follow any subject guidance, skill files, scripts, or configuration under /workspace/input/subject.",
+        "- Candidate source files are mounted at /workspace/input/candidate.",
+        "- Follow any candidate guidance, skill files, scripts, or configuration under /workspace/input/candidate.",
         "- The mutable working directory is /workspace.",
-        "- If the subject declares prepare.command, it has already run and may have copied files into /workspace.",
+        "- If the candidate declares prepare.command, it has already run and may have copied files into /workspace.",
         ...(workload.case?.prompt ? ["Case:", workload.case.prompt, ""] : []),
         "- Public case files are mounted at /workspace/input/case.",
         "- Verifier tests are not present while you run.",
@@ -792,87 +844,89 @@ function buildAgentSubjectPrompt(workload, subject) {
         "- You may write inspection artifacts under /workspace/output.",
     ].join("\n");
 }
-async function writeAgentSubjectRevisionOutput(request, workload, optimizer, options) {
-    if (request.operation !== "optimizer.improve") {
-        throw new Error("Agent subject revision results can only complete optimizer.improve operations.");
+async function writeAgentCandidateRevisionOutput(request, workload, improver, options) {
+    if (request.operation !== "candidate.improve") {
+        throw new Error("Agent improve results can only complete candidate.improve operations.");
     }
-    const traceRoot = path.join(request.paths.output, ".workbench", "internal", "agent-optimizer");
-    const agentResult = await executeBuiltInAgentTurn(options.agentExecutor, {
-        role: "optimizer",
-        provider: optimizer.agent,
-        adapterAuthRoot: options.adapterAuthRoot,
-        adapterAuthRequest: options.adapterAuthRequest,
-        adapterAuthEnv: options.adapterAuthEnv,
-        workspaceRoot: request.paths.workspace,
-        cwd: request.paths.workspace,
-        prompt: buildAgentOptimizerPrompt(workload),
-        traceRoot,
-        jobId: workload.job.id,
-    });
-    const subjectPatch = await createSubjectPatchFromWorkspace({
-        beforeRoot: requiredRequestPath(request.paths.subject, "paths.subject"),
-        afterRoot: request.paths.workspace,
-        edits: workload.optimizer.edits,
-    });
-    const changedSubjectPaths = subjectPatch.fileChanges.filter((filePath) => isSubjectEditPath(filePath, workload.optimizer.edits));
-    if (changedSubjectPaths.length === 0) {
-        throw new Error("Agent improve adapter completed without changing a subject file covered by optimizer edits.");
+    const before = await snapshotEditableCandidateWorkspace(request);
+    const traceRoot = path.join(request.paths.output, ".workbench", "internal", "agent-improver");
+    try {
+        const agentResult = await executeBuiltInAgentTurn(options.agentExecutor, {
+            role: "improver",
+            provider: improver.agent,
+            adapterAuthRoot: options.adapterAuthRoot,
+            adapterAuthRequest: options.adapterAuthRequest,
+            adapterAuthEnv: options.adapterAuthEnv,
+            workspaceRoot: request.paths.workspace,
+            cwd: request.paths.workspace,
+            prompt: buildAgentImproverPrompt(workload),
+            traceRoot,
+            jobId: workload.job.id,
+        });
+        const candidatePatch = await createCandidatePatchFromWorkspace({
+            beforeRoot: before.root,
+            afterRoot: request.paths.workspace,
+            edits: workload.improve.edits,
+        });
+        const changedCandidatePaths = candidatePatch.fileChanges.filter((filePath) => isCandidateEditPath(filePath, workload.improve.edits));
+        if (changedCandidatePaths.length === 0) {
+            throw new Error("Agent improve adapter completed without changing a candidate file covered by improve edits.");
+        }
+        const trace = {
+            path: `.workbench/traces/${workload.job.id}/improver.json`,
+            kind: "text",
+            encoding: "utf8",
+            executable: false,
+            content: `${JSON.stringify({
+                kind: "agent_improver",
+                provider: improver.agent.use,
+                candidateId: workload.candidateId,
+                attemptIndex: workload.attemptIndex,
+                changedPaths: changedCandidatePaths,
+                summary: agentResult.output,
+                metadata: agentResult.metadata,
+            }, null, 2)}\n`,
+        };
+        await writeSurfaceFiles(request.paths.output, [trace, ...agentResult.traceFiles]);
+        const runtime = await importWorkbenchRuntime();
+        const usage = runtime.assignUsageRole("improver", agentResult.usage);
+        await writeWorkbenchAdapterOperationResult(request.paths.output, {
+            protocol: "workbench.adapter-result.v1",
+            operation: "candidate.improve",
+            ok: true,
+            value: {
+                ...candidatePatch,
+                fileChanges: changedCandidatePaths,
+            },
+            ...(agentResult.output ? { summary: agentResult.output } : {}),
+            feedback: {
+                improver: improver.agent.use,
+                changedPaths: changedCandidatePaths,
+                metadata: agentResult.metadata,
+            },
+            ...(usage ? { usage } : {}),
+        });
+    }
+    finally {
+        await before.cleanup();
     }
-    const trace = {
-        path: `.workbench/traces/${workload.job.id}/optimizer.json`,
-        kind: "text",
-        encoding: "utf8",
-        executable: false,
-        content: `${JSON.stringify({
-            kind: "agent_optimizer",
-            provider: optimizer.agent.use,
-            subjectId: workload.subjectId,
-            attemptIndex: workload.attemptIndex,
-            changedPaths: changedSubjectPaths,
-            summary: agentResult.output,
-            metadata: agentResult.metadata,
-        }, null, 2)}\n`,
-    };
-    await writeSurfaceFiles(request.paths.output, [trace, ...agentResult.traceFiles]);
-    const runtime = await importWorkbenchRuntime();
-    const usage = runtime.assignUsageRole("optimizer", agentResult.usage);
-    await writeWorkbenchAdapterOperationResult(request.paths.output, {
-        protocol: "workbench.adapter-result.v1",
-        operation: "optimizer.improve",
-        ok: true,
-        value: {
-            ...subjectPatch,
-            fileChanges: changedSubjectPaths,
-        },
-        ...(agentResult.output ? { summary: agentResult.output } : {}),
-        feedback: {
-            optimizer: optimizer.agent.use,
-            changedPaths: changedSubjectPaths,
-            metadata: agentResult.metadata,
-        },
-        ...(usage ? { usage } : {}),
-    });
 }
-function buildAgentOptimizerPrompt(workload) {
+function buildAgentImproverPrompt(workload) {
     return [
         "Benchmark:",
         workload.benchmark.description || workload.benchmark.name,
         "",
-        "Context:",
-        "- Subject source files are mounted at /workspace/input/subject.",
-        "- Follow any subject guidance, skill files, scripts, or configuration under /workspace/input/subject.",
-        "- The mutable working directory is /workspace.",
-        "- If the subject declares prepare.command, it has already run and may have copied files into /workspace.",
-        "- Prior run traces are mounted at /workspace/input/traces.",
-        "- Use /workspace/input/traces as the source of truth for what happened in prior attempts.",
-        "- Do not mutate /workspace/input.",
+        "Improve the candidate for this benchmark.",
         "",
-        "Editable subject paths:",
-        workload.optimizer.edits.map((entry) => `- ${entry}`).join("\n"),
+        "Candidate files are in the current directory.",
+        "Prior adapter executions are in /workspace/input/traces.",
         "",
-        "Output:",
-        "- Create or mutate editable subject files directly in the current working directory.",
-        "- Include at least one changed subject file covered by the optimizer edits list.",
+        "Editable paths:",
+        workload.improve.edits.map((entry) => `- ${entry}`).join("\n"),
+        "",
+        "Rules:",
+        "- Modify only editable paths.",
+        "- Change at least one editable file.",
     ].join("\n");
 }
 async function writeRubricJudgeResult(request, workload, engine, options = {}) {
@@ -928,9 +982,9 @@ async function writeRubricEvidenceFiles(args) {
     const root = `.workbench/traces/${args.workload.job.id}/engine/rubric`;
     const scorecard = {
         schema: "workbench.engine.rubric.evidence.v1",
-        safeForOptimizer: true,
+        safeForImprover: true,
         jobId: args.workload.job.id,
-        subjectId: args.workload.subjectId,
+        candidateId: args.workload.candidateId,
         attemptIndex: args.workload.attemptIndex,
         sampleIndex: args.workload.sampleIndex,
         caseId: args.workload.caseId,
@@ -957,7 +1011,7 @@ async function writeRubricEvidenceFiles(args) {
         jsonSurfaceFile(`${root}/scorecard.json`, scorecard),
         ...args.criterionRuns.map((run) => jsonSurfaceFile(`${root}/criteria/${safeInternalPathSegment(run.result.criterion_id)}/result.json`, {
             schema: "workbench.engine.rubric.criterion-evidence.v1",
-            safeForOptimizer: true,
+            safeForImprover: true,
             criterion: args.engine.criteria.find((criterion) => criterion.id === run.result.criterion_id) ?? {
                 id: run.result.criterion_id,
             },
@@ -1079,8 +1133,8 @@ function buildRubricCriterionJudgePrompt(workload, engine, criterion) {
         JSON.stringify(criterion, null, 2),
         "",
         "Context:",
-        "- The subject already ran in this same working directory.",
-        "- Subject outputs are available in the current working directory.",
+        "- The candidate already ran in this same working directory.",
+        "- Candidate outputs are available in the current working directory.",
         "- Public case files are mounted at /workspace/input/case.",
         "- Verifier-private files are mounted at /workspace/private/engine when the task provides them.",
         "- Score only from the current working directory, public case files, verifier-private files, and the criterion above.",
@@ -1276,12 +1330,12 @@ function requireWorkloadTask(workload, label) {
         throw new Error(`${label} workload is missing case text.`);
     }
 }
-async function createSubjectPatchFromWorkspace(args) {
+async function createCandidatePatchFromWorkspace(args) {
     const before = new Map((await readSurfaceFilesRecursive(args.beforeRoot))
         .map((file) => [normalizeRelativePath(file.path), file]));
     const changedFiles = (await readSurfaceFilesRecursive(args.afterRoot))
         .map((file) => ({ ...file, path: normalizeRelativePath(file.path) }))
-        .filter((file) => isSubjectEditPath(file.path, args.edits) &&
+        .filter((file) => isCandidateEditPath(file.path, args.edits) &&
         !isRuntimeWorkspacePath(file.path) &&
         !sameSurfaceFile(before.get(file.path), file))
         .sort((left, right) => left.path.localeCompare(right.path));
@@ -1321,7 +1375,7 @@ async function writeSurfaceFiles(root, files) {
         }
     }
 }
-function isSubjectEditPath(filePath, edits) {
+function isCandidateEditPath(filePath, edits) {
     const normalized = normalizeRelativePath(filePath);
     return edits.some((entry) => {
         const editPath = normalizeRelativePath(entry).replace(/\/+$/u, "");

package/dist/manifests.js

@@ -1,4 +1,4 @@
-import { adapterSlot, defineAdapter, defineEngineResolver, defineSubject, defineOptimizer, defineEngineRunner, workbenchAdapterManifestFromDefinition, } from "@workbench-ai/workbench-protocol";
+import { adapterSlot, defineAdapter, defineEngineResolver, defineCandidate, defineImprover, defineEngineRunner, workbenchAdapterManifestFromDefinition, } from "@workbench-ai/workbench-protocol";
 const BUILT_IN_ADAPTER_MANIFESTS = Object.fromEntries(Object.entries({
     workbench: defineAdapter({
         id: "workbench",
@@ -10,8 +10,8 @@ const BUILT_IN_ADAPTER_MANIFESTS = Object.fromEntries(Object.entries({
     }),
     codex: defineAdapter({
         id: "codex",
-        subject: defineSubject(),
-        improve: defineOptimizer(),
+        candidate: defineCandidate(),
+        improve: defineImprover(),
         setup: [
             "npm install --global @openai/codex@0.125.0",
         ],
@@ -24,8 +24,8 @@ const BUILT_IN_ADAPTER_MANIFESTS = Object.fromEntries(Object.entries({
     }),
     claude: defineAdapter({
         id: "claude",
-        subject: defineSubject(),
-        improve: defineOptimizer(),
+        candidate: defineCandidate(),
+        improve: defineImprover(),
         setup: [
             "npm install --global @anthropic-ai/claude-code@2.1.119",
         ],
@@ -57,15 +57,15 @@ const BUILT_IN_ADAPTER_MANIFESTS = Object.fromEntries(Object.entries({
     }),
     command: defineAdapter({
         id: "command",
-        subject: defineSubject(),
+        candidate: defineCandidate(),
         engineRun: defineEngineRunner(),
-        improve: defineOptimizer(),
+        improve: defineImprover(),
     }),
     rubric: defineAdapter({
         id: "rubric",
         engineRun: defineEngineRunner(),
         slots: {
-            judge: adapterSlot("/judge", "subject.run"),
+            judge: adapterSlot("/judge", "candidate.run"),
         },
     }),
     tests: defineAdapter({

package/dist/runtime.js

@@ -12,11 +12,11 @@ export async function importWorkbenchRuntime() {
     return await runtimeModule;
 }
 async function importWorkbenchRuntimeUncached() {
-    const subjects = runtimeImportSubjects();
+    const candidates = runtimeImportCandidates();
     let lastError;
-    for (const subject of subjects) {
+    for (const candidate of candidates) {
         try {
-            return await import(__rewriteRelativeImportExtension(subject));
+            return await import(__rewriteRelativeImportExtension(candidate));
         }
         catch (error) {
             lastError = error;
@@ -24,11 +24,11 @@ async function importWorkbenchRuntimeUncached() {
     }
     throw new Error(`Unable to load @workbench-ai/workbench-core for built-in adapters: ${lastError instanceof Error ? lastError.message : String(lastError)}`);
 }
-function runtimeImportSubjects() {
+function runtimeImportCandidates() {
     return [
         process.env.WORKBENCH_RUNTIME_IMPORT,
         "/app/products/workbench/packages/core/src/index.ts",
         new URL("../../core/src/index.ts", import.meta.url).href,
         "@workbench-ai/workbench-core",
-    ].filter((subject) => typeof subject === "string" && subject.length > 0);
+    ].filter((candidate) => typeof candidate === "string" && candidate.length > 0);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workbench-ai/workbench-built-in-adapters",
-  "version": "0.0.49",
+  "version": "0.0.51",
   "type": "module",
   "repository": {
     "type": "git",
@@ -32,12 +32,12 @@
   ],
   "dependencies": {
     "yaml": "^2.8.2",
-    "@workbench-ai/agent-driver-anthropic-claude-code": "0.0.45",
-    "@workbench-ai/agent-driver": "0.0.45",
     "@workbench-ai/agent-driver-openai-codex": "0.0.45",
-    "@workbench-ai/workbench-protocol": "0.0.49",
-    "@workbench-ai/workbench-core": "0.0.49",
-    "@workbench-ai/workbench-contract": "0.0.49"
+    "@workbench-ai/workbench-contract": "0.0.51",
+    "@workbench-ai/agent-driver": "0.0.45",
+    "@workbench-ai/workbench-core": "0.0.51",
+    "@workbench-ai/agent-driver-anthropic-claude-code": "0.0.45",
+    "@workbench-ai/workbench-protocol": "0.0.51"
   },
   "devDependencies": {
     "@types/node": "^24.3.1",