@workbench-ai/workbench 0.0.53 → 0.0.55

package/dist/benchmark-fingerprint.d.ts

@@ -1,4 +1,6 @@
+import { type WorkbenchProjectStateSource } from "@workbench-ai/workbench-core";
 import { type LocalProjectSource } from "./project-source.js";
 export declare function localBenchmarkFingerprint(project: LocalProjectSource): string;
+export declare function projectStateBenchmarkFingerprint(source: WorkbenchProjectStateSource): string;
 export declare function localCandidateFingerprint(project: LocalProjectSource): string;
 //# sourceMappingURL=benchmark-fingerprint.d.ts.map

package/dist/benchmark-fingerprint.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"benchmark-fingerprint.d.ts","sourceRoot":"","sources":["../src/benchmark-fingerprint.ts"],"names":[],"mappings":"AAOA,OAAO,EAGL,KAAK,kBAAkB,EACxB,MAAM,qBAAqB,CAAC;AAE7B,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,kBAAkB,GAAG,MAAM,CAW7E;AAED,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,kBAAkB,GAAG,MAAM,CAO7E"}
+{"version":3,"file":"benchmark-fingerprint.d.ts","sourceRoot":"","sources":["../src/benchmark-fingerprint.ts"],"names":[],"mappings":"AAAA,OAAO,EAKL,KAAK,2BAA2B,EACjC,MAAM,8BAA8B,CAAC;AAEtC,OAAO,EAGL,KAAK,kBAAkB,EACxB,MAAM,qBAAqB,CAAC;AAE7B,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,kBAAkB,GAAG,MAAM,CAW7E;AAED,wBAAgB,gCAAgC,CAAC,MAAM,EAAE,2BAA2B,GAAG,MAAM,CAU5F;AAED,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,kBAAkB,GAAG,MAAM,CAO7E"}

package/dist/benchmark-fingerprint.js

@@ -12,6 +12,17 @@ export function localBenchmarkFingerprint(project) {
         network: project.spec.environment.network?.egress === "open" ? "on" : "off",
     });
 }
+export function projectStateBenchmarkFingerprint(source) {
+    return workbenchBenchmarkContentFingerprint({
+        sourceYaml: source.source,
+        engineResolveFiles: source.engineResolveFiles,
+        engineResolveBinding: source.engineResolveBinding,
+        adapterFiles: source.adapterFiles,
+        runtimeFiles: source.runtimeFiles,
+        resources: source.resources,
+        network: source.network,
+    });
+}
 export function localCandidateFingerprint(project) {
     return workbenchCandidateContentFingerprint({
         sourceYaml: project.specSource,

package/dist/command-model.js

@@ -140,7 +140,7 @@ const commandHelp = Object.fromEntries(Object.entries({
         "  workbench improve [SOURCE] [--dir DIR] [--from CANDIDATE_ID] [--runs RUN] [--budget N] [--samples N] [--rerun] [--json]",
         "  workbench improve --hosted [SOURCE] [--dir DIR] [--benchmark OWNER/BENCHMARK] [--base CANDIDATE_ID] [--runs RUN] [--budget N] [--samples N] [--rerun] [--watch] [--dry-run] [--json]",
         "",
-        "Ensure a candidate improvement exists for the selected base, run, budget, and samples. Without --hosted, execution writes local records. With --hosted, Workbench starts or reuses hosted work against the configured remote or --benchmark target. Completed work is reused only when base, run configuration, source, adapters, benchmark, budget, and samples match; use --rerun to intentionally spend again.",
+        "Ensure a candidate improvement exists for the selected base, run, budget, and samples. Improve defaults to the evaluated active candidate when it belongs to the current benchmark fingerprint; otherwise it evaluates and uses the authored current candidate. Without --hosted, execution writes local records. With --hosted, Workbench starts or reuses hosted work against the configured remote or --benchmark target. Completed work is reused only when base, run configuration, source, adapters, benchmark, budget, and samples match; use --rerun to intentionally spend again.",
         "",
         "Examples:",
         "  workbench improve --budget 1 --samples 1",

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AA0IA,UAAU,KAAK;IACb,KAAK,EAAE,MAAM,CAAC,cAAc,CAAC;IAC7B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;IAC9B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;CAC/B;AA4BD,UAAU,iBAAiB;CAAG;AA4K9B,wBAAsB,MAAM,CAC1B,IAAI,EAAE,SAAS,MAAM,EAAE,EACvB,EAAE,GAAE,KAIH,EACD,cAAc,GAAE,iBAAsB,GACrC,OAAO,CAAC,MAAM,CAAC,CAmHjB"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AA4IA,UAAU,KAAK;IACb,KAAK,EAAE,MAAM,CAAC,cAAc,CAAC;IAC7B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;IAC9B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;CAC/B;AA4BD,UAAU,iBAAiB;CAAG;AA4K9B,wBAAsB,MAAM,CAC1B,IAAI,EAAE,SAAS,MAAM,EAAE,EACvB,EAAE,GAAE,KAIH,EACD,cAAc,GAAE,iBAAsB,GACrC,OAAO,CAAC,MAAM,CAAC,CAmHjB"}

package/dist/index.js

@@ -5,7 +5,7 @@ import { createRequire } from "node:module";
 import os from "node:os";
 import path from "node:path";
 import { Writable } from "node:stream";
-import { createCandidateFilePreview, createBaselineCandidateJob as createRuntimeBaselineCandidateJob, evaluationScorecardId, evaluationMeanMetrics, executeWorkbenchExecutionJob, engineResolveBindingForSpec, filterOptimizerTraceJobsForCaseIds, filterCandidateSourceFiles, formatWorkbenchCaseSelector, formatWorkbenchSelectionPolicy, workbenchCaseSelectorUsesAllCases, workbenchExecutionPurpose, workbenchRunExecutionFingerprint, createWorkbenchAdapterAuthBundle, createOptimizerTraceInputFiles, DOCKER_SANDBOX_BACKEND, localWorkbenchAdapterAuthStore, materializeWorkbenchRunResult, normalizeSurfaceFiles, planWorkbenchExecutionJobsForPurpose, runWorkbenchExecutionDag, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml, summarizeCandidateFiles, validateWorkbenchRunEnvelope, parseWorkbenchAdapterAuthTarget, workbenchEngineCaseIdsForImproveEvaluation, workbenchEngineCaseIdsForSelector, workbenchImproveOptimizeSelector, workbenchImproveSelectionPolicy, workbenchProjectSourceFingerprint, workbenchRuntimeBundleFingerprint, } from "@workbench-ai/workbench-core";
+import { createCandidateFilePreview, createBaselineCandidateJob as createRuntimeBaselineCandidateJob, evaluationScorecardId, evaluationMeanMetrics, executeWorkbenchExecutionJob, engineResolveBindingForSpec, filterOptimizerTraceJobsForCaseIds, filterCandidateSourceFiles, formatWorkbenchCaseSelector, formatWorkbenchSelectionPolicy, workbenchCaseSelectorUsesAllCases, workbenchExecutionPurpose, workbenchRunExecutionFingerprint, createWorkbenchAdapterAuthBundle, createOptimizerTraceInputFiles, DOCKER_SANDBOX_BACKEND, localWorkbenchAdapterAuthStore, materializeWorkbenchRunResult, normalizeSurfaceFiles, planWorkbenchExecutionJobsForPurpose, runWorkbenchExecutionDag, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml, summarizeCandidateFiles, validateWorkbenchRunEnvelope, parseWorkbenchAdapterAuthTarget, workbenchEngineCaseIdsForImproveEvaluation, workbenchEngineCaseIdsForSelector, workbenchImproveOptimizeSelector, workbenchImproveSelectionPolicy, workbenchProjectSourceFingerprint, workbenchRuntimeBundleFingerprint, workbenchRuntimeExplicitActiveId, } from "@workbench-ai/workbench-core";
 import { assertWorkbenchAdapterOperationResultOk, collectWorkbenchAdapterAuthRequirements, normalizeWorkbenchAdapterOperationRequest, readWorkbenchAdapterOperationResult, workbenchAdapterOperationCommand, workbenchAdapterOperationResultPath, withDefaultWorkbenchAdapterAuthProfiles as applyDefaultWorkbenchAdapterAuthProfiles, } from "@workbench-ai/workbench-protocol";
 import { builtinLocalTraceAdapter, builtinLocalTraceAdapters, sortLocalTraceRefs, } from "@workbench-ai/workbench-built-in-adapters/local-traces";
 import { commandUsage, HOSTED_WATCH_LIFECYCLE_NOTE, LOCAL_DEV_OPEN_LIFECYCLE_NOTE, rootUsage, } from "./command-model.js";
@@ -16,7 +16,7 @@ import { createAdapterCommandEnv } from "./adapter-command-env.js";
 import { loadLocalArchive, loadLocalArchiveIndex, exportLocalRuntimeBundle, importLocalRuntimeBundle, runtimeBundleStats, materializeCandidateRoot, readLocalCandidate, readLocalCandidateFiles, readLocalJobs, saveLocalArchive, saveLocalJobs, setLocalActive, upsertLocalRun, upsertLocalCandidate, upsertLocalEvaluation, } from "./local-archive.js";
 import { WorkspaceSnapshotError, } from "./workspace-snapshot.js";
 import { hostedEngineResolveFiles, readLocalProjectSource, WORKBENCH_BENCHMARK_FILE, } from "./project-source.js";
-import { localBenchmarkFingerprint, localCandidateFingerprint, } from "./benchmark-fingerprint.js";
+import { localBenchmarkFingerprint, localCandidateFingerprint, projectStateBenchmarkFingerprint, } from "./benchmark-fingerprint.js";
 const require = createRequire(import.meta.url);
 function getCliVersion() {
     const manifest = require("../package.json");
@@ -993,35 +993,30 @@ async function localRun(argv, io, runtimeOptions) {
 async function ensureLocalImproveBaseCandidate(args) {
     let snapshot = await loadLocalArchive(args.workspace);
     const explicitBase = asOptionalString(args.parsed.flags.from);
-    const benchmarkFingerprint = await readLocalBenchmarkFingerprint(args.workspace);
+    const benchmarkFingerprint = localBenchmarkFingerprint(args.projectSource);
+    const baseCandidateArgs = {
+        workspace: args.workspace,
+        benchmarkFingerprint,
+        projectSource: args.projectSource,
+        samples: args.samples,
+        rerun: args.parsed.flags.rerun === true,
+        io: args.io,
+        runtimeOptions: args.runtimeOptions,
+    };
     if (explicitBase) {
-        let candidate = readLocalCandidate(snapshot, explicitBase);
-        if (candidate.benchmarkFingerprint !== benchmarkFingerprint) {
-            throw new UsageError(`Base candidate ${explicitBase} belongs to benchmark ${candidate.benchmarkFingerprint}, not ${benchmarkFingerprint}.`);
-        }
-        if (!candidate.candidateFingerprint) {
-            throw new UsageError(`Base candidate ${explicitBase} is missing a candidate fingerprint.`);
-        }
-        if (candidate.status !== "evaluated" && !candidate.eval) {
-            const code = await localEvaluateCandidate([
-                "--dir",
-                args.workspace,
-                "--candidate",
-                explicitBase,
-                "--runs",
-                args.projectSource.spec.candidate.selectedRunId,
-                "--samples",
-                String(args.samples),
-                ...(args.parsed.flags.rerun === true ? ["--rerun"] : []),
-                "--json",
-            ], createSilentIo(args.io), args.runtimeOptions);
-            if (code !== 0) {
-                throw new UsageError(`Base candidate ${explicitBase} eval failed; improve was not started.`);
-            }
-            snapshot = await loadLocalArchive(args.workspace);
-            candidate = readLocalCandidate(snapshot, explicitBase);
+        return await ensureEvaluatedLocalImproveBaseCandidate({
+            ...baseCandidateArgs,
+            candidateId: explicitBase,
+        });
+    }
+    if (snapshot.activeId) {
+        const activeCandidate = readLocalCandidate(snapshot, snapshot.activeId);
+        if (activeCandidate.benchmarkFingerprint === benchmarkFingerprint) {
+            return await ensureEvaluatedLocalImproveBaseCandidate({
+                ...baseCandidateArgs,
+                candidateId: activeCandidate.id,
+            });
         }
-        return candidate;
     }
     const candidateFingerprint = localCandidateFingerprint(args.projectSource);
     const existing = snapshot.candidates.find((candidate) => candidate.benchmarkFingerprint === benchmarkFingerprint &&
@@ -1063,6 +1058,36 @@ async function ensureLocalImproveBaseCandidate(args) {
     }
     return evaluated;
 }
+async function ensureEvaluatedLocalImproveBaseCandidate(args) {
+    let snapshot = await loadLocalArchive(args.workspace);
+    let candidate = readLocalCandidate(snapshot, args.candidateId);
+    if (candidate.benchmarkFingerprint !== args.benchmarkFingerprint) {
+        throw new UsageError(`Base candidate ${args.candidateId} belongs to benchmark ${candidate.benchmarkFingerprint}, not ${args.benchmarkFingerprint}.`);
+    }
+    if (!candidate.candidateFingerprint) {
+        throw new UsageError(`Base candidate ${args.candidateId} is missing a candidate fingerprint.`);
+    }
+    if (candidate.status === "evaluated" || candidate.eval) {
+        return candidate;
+    }
+    const code = await localEvaluateCandidate([
+        "--dir",
+        args.workspace,
+        "--candidate",
+        args.candidateId,
+        "--runs",
+        args.projectSource.spec.candidate.selectedRunId,
+        "--samples",
+        String(args.samples),
+        ...(args.rerun ? ["--rerun"] : []),
+        "--json",
+    ], createSilentIo(args.io), args.runtimeOptions);
+    if (code !== 0) {
+        throw new UsageError(`Base candidate ${args.candidateId} eval failed; improve was not started.`);
+    }
+    snapshot = await loadLocalArchive(args.workspace);
+    return readLocalCandidate(snapshot, args.candidateId);
+}
 function createSilentIo(io) {
     const sink = new class extends Writable {
         _write(_chunk, _encoding, callback) {
@@ -2922,6 +2947,7 @@ async function pushBenchmark(argv, io) {
     const runtime = await exportLocalRuntimeBundle(dir, {
         currentBenchmarkFingerprint: localBenchmarkFingerprint(source),
     });
+    const localRuntimeFingerprint = workbenchRuntimeBundleFingerprint(runtime);
     const state = localProjectState({
         source,
         runtime,
@@ -2941,7 +2967,7 @@ async function pushBenchmark(argv, io) {
                 sourceFileCount: sourceFileCount(source),
                 runtime: runtimeBundleStats(runtime),
                 sourceFingerprint: state.source.fingerprint,
-                runtimeFingerprint: state.base.runtimeFingerprint,
+                runtimeFingerprint: localRuntimeFingerprint,
             }, parsed, io, () => `Would push benchmark ${source.spec.name}.`);
             return 0;
         }
@@ -2990,7 +3016,7 @@ async function pushBenchmark(argv, io) {
             sourceFileCount: sourceFileCount(source),
             runtime: runtimeBundleStats(runtime),
             sourceFingerprint: state.source.fingerprint,
-            runtimeFingerprint: state.base.runtimeFingerprint,
+            runtimeFingerprint: localRuntimeFingerprint,
         }, parsed, io, () => `Would push ${sourceFileCount(source)} source file(s) and runtime history to ${origin.remote}.`);
         return 0;
     }
@@ -3081,7 +3107,7 @@ async function cloneProject(argv, io) {
             ref,
             outputDir,
             fileCount: state.source.files.length,
-            runtime: runtimeBundleStats(state.runtime),
+            runtime: projectStateRuntimeStats(state),
             sourceFingerprint: state.source.fingerprint ?? state.base.sourceFingerprint ?? null,
             runtimeFingerprint: state.base.runtimeFingerprint ?? null,
         }, parsed, io, () => `Would clone ${formatBenchmarkRef(ref)} to ${outputDir}.`);
@@ -3121,7 +3147,7 @@ async function pullProject(argv, io) {
             dryRun: true,
             dir,
             fileCount: state.source.files.length,
-            runtime: runtimeBundleStats(state.runtime),
+            runtime: projectStateRuntimeStats(state),
             sourceFingerprint: state.source.fingerprint ?? state.base.sourceFingerprint ?? null,
             runtimeFingerprint: state.base.runtimeFingerprint ?? null,
         }, parsed, io, () => `Would pull ${state.source.files.length} source file(s) and runtime history into ${dir}.`);
@@ -3902,6 +3928,18 @@ function localProjectState(args) {
         runtime,
     };
 }
+function projectStateRuntimeStats(state) {
+    const activeId = workbenchRuntimeExplicitActiveId({
+        candidates: state.runtime.candidates,
+        runs: state.runtime.runs,
+        preferredActiveId: state.runtime.activeId ?? null,
+        benchmarkFingerprint: projectStateBenchmarkFingerprint(state.source),
+    });
+    return runtimeBundleStats({
+        ...state.runtime,
+        activeId,
+    });
+}
 function localCandidateRecord(candidate) {
     return {
         ...candidate,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workbench-ai/workbench",
-  "version": "0.0.53",
+  "version": "0.0.55",
   "type": "module",
   "repository": {
     "type": "git",
@@ -21,9 +21,9 @@
   ],
   "dependencies": {
     "yaml": "^2.8.2",
-    "@workbench-ai/workbench-built-in-adapters": "0.0.53",
-    "@workbench-ai/workbench-core": "0.0.53",
-    "@workbench-ai/workbench-protocol": "0.0.53"
+    "@workbench-ai/workbench-protocol": "0.0.55",
+    "@workbench-ai/workbench-core": "0.0.55",
+    "@workbench-ai/workbench-built-in-adapters": "0.0.55"
   },
   "devDependencies": {
     "@tailwindcss/postcss": "^4.2.2",