@workbench-ai/workbench 0.0.78 → 0.0.80

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAiEA,MAAM,WAAW,KAAK;IACpB,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;IAC9B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;CAC/B;AAuTD,wBAAsB,MAAM,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,EAAE,EAAE,GAAE,KAGzD,GAAG,OAAO,CAAC,MAAM,CAAC,CAoMlB"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAkEA,MAAM,WAAW,KAAK;IACpB,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;IAC9B,MAAM,EAAE,MAAM,CAAC,cAAc,CAAC;CAC/B;AAuTD,wBAAsB,MAAM,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,EAAE,EAAE,GAAE,KAGzD,GAAG,OAAO,CAAC,MAAM,CAAC,CAoMlB"}

package/dist/index.js

@@ -4,7 +4,7 @@ import { createRequire } from "node:module";
 import os from "node:os";
 import path from "node:path";
 import { gzipSync } from "node:zlib";
-import { addWorkbenchRemote, addWorkbenchAgent, compareWorkbench, createWorkbenchInspectionSnapshot, createWorkbenchAdapterAuthBundle, createWorkbenchReadOnlyInspectionSnapshot, diffWorkbenchVersions, evalWorkbenchSkill, improveWorkbenchSkill, initWorkbenchSkill, listWorkbenchAgents, listWorkbenchVersions, localWorkbenchAdapterAuthStore, parseWorkbenchAdapterAuthTarget, prepareWorkbenchCloudEvalRequest, prepareWorkbenchCloudImproveRequest, publishWorkbenchVersion, removeWorkbenchAgent, showWorkbenchRef, switchWorkbenchVersion, syncWorkbenchRemote, workbenchJobEvidenceForSnapshot, workbenchStatusSnapshot, WorkbenchCodedError, WorkbenchUserError, } from "@workbench-ai/workbench-core";
+import { addWorkbenchRemote, addWorkbenchAgent, compareWorkbench, createWorkbenchVersionRuntimeSnapshot, createWorkbenchInspectionSnapshot, createWorkbenchAdapterAuthBundle, createWorkbenchReadOnlyInspectionSnapshot, diffWorkbenchVersions, evalWorkbenchSkill, improveWorkbenchSkill, initWorkbenchSkill, listWorkbenchAgents, listWorkbenchVersions, localWorkbenchAdapterAuthStore, parseWorkbenchAdapterAuthTarget, prepareWorkbenchCloudEvalRequest, prepareWorkbenchCloudImproveRequest, publishWorkbenchVersion, removeWorkbenchAgent, showWorkbenchRef, switchWorkbenchVersion, syncWorkbenchRemote, workbenchJobEvidenceForSnapshot, workbenchStatusSnapshot, WorkbenchCodedError, WorkbenchUserError, } from "@workbench-ai/workbench-core";
 import { normalizeWorkbenchSkillName } from "@workbench-ai/workbench-contract";
 import { emitError, emitResult } from "./output.js";
 import { fanOutSkill, manualFanOutCommand } from "./fanout.js";
@@ -582,13 +582,15 @@ async function handleShow(parsed, io) {
     if (artifact) {
         return output(fileListing("artifact", artifact.id, artifact.files), parsed, io, () => formatFileListing("artifact", artifact.id, artifact.files));
     }
-    const details = evidenceDetailsForRunOrJob(snapshot, objectRef);
-    const evidenceFiles = evidenceFilesForRunOrJob(snapshot, objectRef);
-    if (details.length > 0 || evidenceFiles.length > 0) {
+    const selection = runOrJobEvidenceSelection(snapshot, objectRef);
+    const details = evidenceDetailsForSelection(snapshot, selection);
+    const evidenceFiles = evidenceFilesForSelection(snapshot, selection);
+    if (selection.run || selection.jobs.length > 0 || details.length > 0 || evidenceFiles.length > 0) {
         return output({
+            jobs: selection.jobs.map(jobEvidenceSummary),
             details: details.map(evidenceDetailSummary),
             files: evidenceFiles.map(fileSummary),
-        }, parsed, io, () => formatRunOrJobEvidence(details, evidenceFiles));
+        }, parsed, io, () => formatRunOrJobEvidence(selection.jobs, details, evidenceFiles));
     }
     const value = await showWorkbenchRef(ref, core);
     return output(value, parsed, io, () => formatShow(value));
@@ -723,6 +725,7 @@ const API_REQUEST_MAX_ATTEMPTS = 3;
 const API_REQUEST_GZIP_THRESHOLD_BYTES = 1024 * 1024;
 const CLOUD_RUN_TIMEOUT_MS = 30 * 60 * 1000;
 const CLOUD_RUN_POLL_INTERVAL_MS = 3000;
+const LOGIN_WAIT_TIMEOUT_SECONDS = 120;
 async function handleLogin(parsed, io) {
     const provider = optionalPositional(parsed, 1);
     if (provider) {
@@ -739,16 +742,23 @@ async function handleLogin(parsed, io) {
     }
     if (parsed.flags["start-only"] === true && parsed.flags.wait === true) {
         throw new WorkbenchCodedError("usage", "workbench login accepts only one of --start-only or --wait.", {
-            remediation: "Run workbench login --start-only or workbench login --wait.",
+            remediation: `Run workbench login --start-only or workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}.`,
             exitCode: 2,
         });
     }
-    const startOnly = parsed.flags["start-only"] === true;
+    const startOnly = parsed.flags["start-only"] === true ||
+        (parsed.flags["no-open"] === true && parsed.flags.wait !== true && parsed.flags.timeout === undefined);
     const waitOnly = parsed.flags.wait === true;
     const timeoutSeconds = intFlag(parsed, "timeout");
     if (startOnly && timeoutSeconds !== undefined) {
         throw new WorkbenchCodedError("usage", "workbench login --timeout only applies with --wait.", {
-            remediation: "Run workbench login --start-only, then workbench login --wait.",
+            remediation: `Run workbench login --start-only, then workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}.`,
+            exitCode: 2,
+        });
+    }
+    if (waitOnly && timeoutSeconds === undefined) {
+        throw new WorkbenchCodedError("usage", "workbench login --wait requires --timeout N.", {
+            remediation: `Run workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}.`,
             exitCode: 2,
         });
     }
@@ -773,8 +783,8 @@ async function handleLogin(parsed, io) {
             verificationUriComplete: record.verification_uri_complete,
             userCode: record.user_code,
             expiresAt: record.expiresAt,
-            resume: "workbench login --wait",
-        }, parsed, io, () => `Open ${record.verification_uri_complete}\nCode: ${record.user_code}\nResume: workbench login --wait`);
+            resume: `workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}`,
+        }, parsed, io, () => `Open ${record.verification_uri_complete}\nCode: ${record.user_code}\nResume: workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}`);
     }
     await writePendingDeviceAuthorization(record);
     if (freshAuthorization && !parsed.flags.json) {
@@ -1012,6 +1022,9 @@ function installNextCommand(fanout) {
 }
 function formatInstallOutcome(result, dryRun) {
     if (dryRun) {
+        if (result.result === "unchanged") {
+            return `Already installed ${result.directoryName} at ${result.destination} (unchanged; dry run made no changes).`;
+        }
         return `Would install ${result.directoryName} to ${result.destination} (${formatFileCount(result.filesCopied)}).`;
     }
     if (result.result === "unchanged") {
@@ -1075,7 +1088,7 @@ function withTimeout(promise, timeoutMs) {
 }
 async function startCloudExecution(command, parsed, io) {
     const root = dirFlag(parsed) ?? process.cwd();
-    const showProgress = parsed.flags.json !== true;
+    const showProgress = true;
     const interrupt = createCloudInterruptController(command, io, showProgress);
     try {
         writeCloudProgress(io, `workbench cloud: preparing hosted ${command}.`, showProgress);
@@ -1111,6 +1124,14 @@ async function startCloudExecution(command, parsed, io) {
                 samples: intFlag(parsed, "samples"),
                 budget: intFlag(parsed, "budget"),
             }));
+        writeCloudProgress(io, "workbench cloud: checking provider auth.", showProgress);
+        await cloudPreScheduleStep(command, interrupt, preflightCloudAdapterAuth({
+            root,
+            versionId: request.versionId,
+            parsed,
+            baseUrl: source.baseUrl,
+            authToken: token,
+        }));
         writeCloudProgress(io, "workbench cloud: syncing source to cloud.", showProgress);
         const syncBefore = await cloudPreScheduleStep(command, interrupt, syncWorkbenchRemote({ ...core, remote: remote.name }));
         writeCloudProgress(io, `workbench cloud: scheduling hosted ${command}.`, showProgress);
@@ -1209,6 +1230,79 @@ function cloudCanceledBeforeRunIdError(command) {
         exitCode: 130,
     });
 }
+async function preflightCloudAdapterAuth(input) {
+    const snapshot = await createWorkbenchReadOnlyInspectionSnapshot({ dir: input.root, authToken: input.authToken });
+    const version = snapshotVersionByRef(snapshot, input.versionId);
+    if (!version) {
+        throw new WorkbenchCodedError("version_not_found", `Version not found: ${input.versionId}`, {
+            remediation: "Run workbench status.",
+            subject: { versionId: input.versionId },
+            exitCode: 1,
+        });
+    }
+    const runtime = await createWorkbenchVersionRuntimeSnapshot(version, {
+        skill: stringFlag(input.parsed, "skills"),
+        agent: stringFlag(input.parsed, "agents"),
+        authToken: input.authToken,
+    });
+    const targets = uniqueAdapterAuthTargets(runtime.selectedAgents.flatMap(cloudAdapterAuthTargetsForAgent));
+    if (targets.length === 0) {
+        return;
+    }
+    const statuses = await fetchCloudAdapterAuthStatuses(input.baseUrl);
+    const missing = targets.find((target) => !statuses.some((status) => adapterAuthStatusMatchesTarget(status, target)));
+    if (!missing) {
+        return;
+    }
+    throw new WorkbenchCodedError("adapter_auth_required", `${formatCloudAdapterAuthTarget(missing)} disconnected. Run workbench login ${missing.adapterId}.`, {
+        remediation: `Run workbench login ${missing.adapterId}.`,
+        subject: {
+            adapterId: missing.adapterId,
+            profile: missing.profile,
+            ...(missing.slot ? { slot: missing.slot } : {}),
+        },
+        exitCode: 1,
+    });
+}
+function cloudAdapterAuthTargetsForAgent(agent) {
+    const adapterId = agent.adapter.trim().toLowerCase();
+    if (adapterId !== "codex" && adapterId !== "claude") {
+        return [];
+    }
+    const auth = agent.config.auth;
+    if (typeof auth === "string" && auth.trim()) {
+        return [{ adapterId, profile: auth.trim() }];
+    }
+    if (auth && typeof auth === "object" && !Array.isArray(auth)) {
+        return Object.entries(auth)
+            .filter((entry) => typeof entry[1] === "string" && entry[1].trim().length > 0)
+            .map(([slot, profile]) => ({ adapterId, slot, profile: profile.trim() }));
+    }
+    return [{ adapterId, profile: "default" }];
+}
+function uniqueAdapterAuthTargets(targets) {
+    const byKey = new Map();
+    for (const target of targets) {
+        byKey.set(adapterAuthTargetKey(target), target);
+    }
+    return [...byKey.values()].sort((left, right) => adapterAuthTargetKey(left).localeCompare(adapterAuthTargetKey(right)));
+}
+async function fetchCloudAdapterAuthStatuses(baseUrl) {
+    const response = await apiRequest("/api/workbench/auth/adapters", {}, baseUrl);
+    return response.adapters ?? [];
+}
+function adapterAuthStatusMatchesTarget(status, target) {
+    return status.status === "connected" &&
+        status.adapterId === target.adapterId &&
+        status.profile === target.profile &&
+        (status.slot ?? undefined) === (target.slot ?? undefined);
+}
+function adapterAuthTargetKey(target) {
+    return `${target.adapterId}/${target.slot ?? "_"}/${target.profile}`;
+}
+function formatCloudAdapterAuthTarget(target) {
+    return `${target.adapterId}${target.slot ? `/${target.slot}` : ""}`;
+}
 async function waitForCloudRuns(input) {
     const runIds = input.runs
         .map((run) => run.id)
@@ -1726,7 +1820,7 @@ async function pollDeviceToken(baseUrl, authorization, timeoutSeconds) {
     }
     throw new WorkbenchCodedError("login_pending", "Device login is still waiting for browser authorization.", {
         retryable: true,
-        remediation: "Authorize the device in the browser, then run workbench login --wait.",
+        remediation: `Authorize the device in the browser, then run workbench login --wait --timeout ${LOGIN_WAIT_TIMEOUT_SECONDS}.`,
         subject: {
             retryAfterSeconds: Math.max(1, Math.ceil(intervalMs / 1000)),
             verificationUri: authorization.verification_uri,
@@ -2750,11 +2844,6 @@ function snapshotHasWorkflowCase(snapshot) {
         /^\.workbench\/cases\/[^/]+\/case\.ya?ml$/u.test(file.path)) ?? [];
     return caseFiles.some((file) => file.kind === "text" && !/\n\s*smoke:\s*true(?:\s|$)/u.test(`\n${file.content}`));
 }
-function installHandleFromStatusRemote(remote) {
-    const publicationUrl = remote.publication.status === "published" ? remote.publication.installUrl : undefined;
-    const source = parseWorkbenchInstallSource(publicationUrl ?? remote.url);
-    return source ? `${source.owner}/${source.skill}` : publicationUrl ?? remote.url;
-}
 async function statusWithCausalNext(status, auth, core, machine) {
     if (!status.project.initialized) {
         return {
@@ -2816,12 +2905,6 @@ async function statusWithCausalNext(status, auth, core, machine) {
         }
         return { ...status, next: "workbench publish" };
     }
-    const publishedCloudRemote = status.remotes.find((remote) => remote.kind === "workbench-cloud" &&
-        remote.publication.status === "published" &&
-        Boolean(remote.publication.installUrl));
-    if (publishedCloudRemote) {
-        return { ...status, next: `workbench install ${installHandleFromStatusRemote(publishedCloudRemote)}` };
-    }
     return {
         ...status,
         next: null,
@@ -2880,6 +2963,16 @@ function displayRef(id) {
 function shortenCommandRefs(command) {
     return command.replace(/\b(?:v_[0-9a-f]{8,}|(?:run|job|trace|artifact)_[a-z0-9_-]+)/giu, (match) => displayRef(match));
 }
+function displayCandidateRefs(ids) {
+    const uniqueIds = [...ids];
+    for (let length = 8; length <= 32; length += 1) {
+        const refs = uniqueIds.map((id) => id.length > length ? id.slice(0, length) : id);
+        if (new Set(refs).size === refs.length) {
+            return refs;
+        }
+    }
+    return uniqueIds;
+}
 function snapshotVersionByRef(snapshot, ref) {
     const requested = ref.trim();
     const normalized = requested === "current" ? snapshot.refs.current ?? "" : requested;
@@ -2888,7 +2981,7 @@ function snapshotVersionByRef(snapshot, ref) {
     }
     const candidates = snapshot.versions.filter((version) => snapshotVersionRefMatches(version, normalized));
     if (candidates.length > 1) {
-        throw new WorkbenchCodedError("ref_ambiguous", `Version ref is ambiguous: ${ref}. Candidates: ${candidates.map((version) => displayRef(version.id)).join(", ")}.`, {
+        throw new WorkbenchCodedError("ref_ambiguous", `Version ref is ambiguous: ${ref}. Candidates: ${displayCandidateRefs(candidates.map((version) => version.id)).join(", ")}.`, {
             subject: { ref, candidates: candidates.map((version) => version.id) },
             exitCode: 2,
         });
@@ -2911,7 +3004,7 @@ function snapshotObjectByRef(entries, ref, kind) {
     }
     const candidates = entries.filter((entry) => objectRefMatches(entry.id, normalized));
     if (candidates.length > 1) {
-        throw new WorkbenchCodedError("ref_ambiguous", `${capitalize(kind)} ref is ambiguous: ${ref}. Candidates: ${candidates.map((entry) => displayRef(entry.id)).slice(0, 8).join(", ")}.`, {
+        throw new WorkbenchCodedError("ref_ambiguous", `${capitalize(kind)} ref is ambiguous: ${ref}. Candidates: ${displayCandidateRefs(candidates.map((entry) => entry.id)).slice(0, 8).join(", ")}.`, {
             subject: { ref, candidates: candidates.map((entry) => entry.id).slice(0, 20) },
             exitCode: 2,
         });
@@ -2932,7 +3025,7 @@ function runOrJobEvidenceSelection(snapshot, ref) {
     const run = snapshotObjectByRef(snapshot.runs, ref, "run");
     const job = snapshotObjectByRef(snapshot.jobs, ref, "job");
     if (run && job) {
-        throw new WorkbenchCodedError("ref_ambiguous", `Run/job ref is ambiguous: ${ref}. Candidates: ${displayRef(run.id)}, ${displayRef(job.id)}.`, {
+        throw new WorkbenchCodedError("ref_ambiguous", `Run/job ref is ambiguous: ${ref}. Candidates: ${displayCandidateRefs([run.id, job.id]).join(", ")}.`, {
             subject: { ref, candidates: [run.id, job.id] },
             exitCode: 2,
         });
@@ -2947,6 +3040,9 @@ function runOrJobEvidenceSelection(snapshot, ref) {
 }
 function evidenceFilesForRunOrJob(snapshot, ref) {
     const selection = runOrJobEvidenceSelection(snapshot, ref);
+    return evidenceFilesForSelection(snapshot, selection);
+}
+function evidenceFilesForSelection(snapshot, selection) {
     if (!selection.run && selection.jobs.length === 0) {
         return [];
     }
@@ -3028,10 +3124,32 @@ function isUserFacingTraceEvidenceFile(file) {
 function evidencePathSegment(value) {
     return value.replace(/[^A-Za-z0-9._-]+/gu, "-") || "_";
 }
-function formatRunOrJobEvidence(details, files) {
+function formatRunOrJobEvidence(jobs, details, files) {
+    const jobLines = jobs.length > 0 ? ["Jobs:", ...jobs.map(formatJobEvidenceSummary)] : [];
     const detailLines = details.map(formatTraceDetail).filter(Boolean);
     const fileLines = files.length > 0 ? ["Files:", ...files.map((file) => file.path)] : [];
-    return [...detailLines, ...fileLines].join("\n") || "No evidence.";
+    return [...jobLines, ...detailLines, ...fileLines].join("\n") || "No evidence.";
+}
+function jobEvidenceSummary(job) {
+    return {
+        id: job.id,
+        runId: job.runId,
+        caseId: job.caseId,
+        sample: job.sample,
+        status: job.status,
+        ...(job.score !== undefined ? { score: job.score } : {}),
+        ...(job.error ? { error: job.error } : {}),
+    };
+}
+function formatJobEvidenceSummary(job) {
+    return [
+        displayRef(job.id),
+        `case=${job.caseId}`,
+        `sample=${job.sample}`,
+        job.status,
+        job.score !== undefined ? `score=${job.score.toFixed(3)}` : undefined,
+        job.error ? `error=${singleLine(job.error)}` : undefined,
+    ].filter(Boolean).join("\t");
 }
 function evidenceDetailSummary(detail) {
     return {
@@ -3107,6 +3225,9 @@ async function fileForRunOrJobRef(core, objectRef, requestedPath) {
 }
 function evidenceDetailsForRunOrJob(snapshot, ref) {
     const selection = runOrJobEvidenceSelection(snapshot, ref);
+    return evidenceDetailsForSelection(snapshot, selection);
+}
+function evidenceDetailsForSelection(snapshot, selection) {
     return selection.jobs.flatMap((entry) => {
         const detail = workbenchJobEvidenceForSnapshot(snapshot, {
             runId: entry.runId,
@@ -3254,7 +3375,21 @@ async function evalSuccessNextCommand(core, runs) {
         return "edit .workbench/cases, then run workbench eval";
     }
     const snapshot = await createWorkbenchReadOnlyInspectionSnapshot(core);
-    return snapshotHasWorkflowCase(snapshot) ? "workbench publish" : "edit .workbench/cases, then run workbench eval";
+    if (!snapshotHasWorkflowCase(snapshot)) {
+        return "edit .workbench/cases, then run workbench eval";
+    }
+    const auth = await workbenchCliAuthStatus();
+    if (auth.workbenchCloud.status !== "authenticated") {
+        return "workbench login";
+    }
+    const status = await workbenchStatusSnapshot(core);
+    return statusHasPublishedCurrentCloudSource(status) ? null : "workbench publish";
+}
+function statusHasPublishedCurrentCloudSource(status) {
+    const currentVersionId = status.project.currentVersionId;
+    return Boolean(currentVersionId && status.remotes.some((remote) => remote.kind === "workbench-cloud" &&
+        remote.publication.status === "published" &&
+        remote.publication.versionId === currentVersionId));
 }
 function formatStatusSnapshot(status) {
     const lines = [

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workbench-ai/workbench",
-  "version": "0.0.78",
+  "version": "0.0.80",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/workbench-ai/workbench.git",
@@ -22,10 +22,10 @@
   "dependencies": {
     "skills": "1.5.11",
     "yaml": "^2.8.2",
-    "@workbench-ai/workbench-built-in-adapters": "0.0.78",
-    "@workbench-ai/workbench-protocol": "0.0.78",
-    "@workbench-ai/workbench-contract": "0.0.78",
-    "@workbench-ai/workbench-core": "0.0.78"
+    "@workbench-ai/workbench-built-in-adapters": "0.0.80",
+    "@workbench-ai/workbench-protocol": "0.0.80",
+    "@workbench-ai/workbench-contract": "0.0.80",
+    "@workbench-ai/workbench-core": "0.0.80"
   },
   "devDependencies": {
     "@tailwindcss/postcss": "^4.2.2",
@@ -36,7 +36,7 @@
     "react-dom": "^19.2.0",
     "typescript": "^5.9.2",
     "vitest": "^3.2.4",
-    "@workbench-ai/workbench-ui": "0.0.78"
+    "@workbench-ai/workbench-ui": "0.0.80"
   },
   "scripts": {
     "build": "rm -rf dist && tsc -p tsconfig.json && chmod 755 dist/workbench.js && node ./scripts/build-dev-open-assets.mjs",