@workbench-ai/workbench 0.0.54 → 0.0.55

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.js

@@ -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;
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workbench-ai/workbench",
-  "version": "0.0.54",
+  "version": "0.0.55",
   "type": "module",
   "repository": {
     "type": "git",
@@ -21,9 +21,9 @@
   ],
   "dependencies": {
     "yaml": "^2.8.2",
-    "@workbench-ai/workbench-built-in-adapters": "0.0.54",
-    "@workbench-ai/workbench-core": "0.0.54",
-    "@workbench-ai/workbench-protocol": "0.0.54"
+    "@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",