@pome-sh/cli 0.39.1 → 0.40.0

package/dist/build-info.json

@@ -1,6 +1,6 @@
 {
   "package": "pome-sh",
-  "version": "0.39.1",
-  "git_sha": "ad57e84671dd2e6a2fb0dc3b66a78d97b95ed1c7",
-  "build_time": "2026-08-30T05:08:05.900Z"
+  "version": "0.40.0",
+  "git_sha": "18ff79b115d837fa3081e5925664bd24f5843c8c",
+  "build_time": "2026-08-30T05:50:34.377Z"
 }

package/dist/{chunk-3NQUFSIX.js → chunk-G7252TQ3.js}

@@ -440,7 +440,7 @@ function mapHttpError(status, json) {
   }
   if (status === 404) {
     return new HostedOrchError(
-      err?.message ?? "POST /v1/agents is not available at this API URL. Upgrade pome-cloud or check that --api-url/POME_API_URL points at a version that supports agent registration."
+      err?.message ?? "POST /v1/agents returned 404 \u2014 check that --api-url/POME_API_URL points at a Pome control plane."
     );
   }
   return new HostedOrchError(err?.message ?? `POST /v1/agents \u2192 HTTP ${status}`);

package/dist/{runTrialGroup-YACZ2WTV.js → runTrialGroup-I6FCP2L5.js}

@@ -1,5 +1,5 @@
 import { newGroupId } from './chunk-JNXZBK3O.js';
-import { runTaskHosted, resolveRunAgentIdentity } from './chunk-3NQUFSIX.js';
+import { runTaskHosted, resolveRunAgentIdentity } from './chunk-G7252TQ3.js';
 import { createHostedClient, parseTaskFile, outcomeOf, isNarrated, criterionPhrase, narratorReadingLines } from './chunk-BS47AE55.js';
 import './chunk-VPXIGCZW.js';
 import './chunk-NW7HGA2K.js';

package/dist/src/cli/main.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { DEFAULT_CONTROL_PLANE_URL, DEFAULT_DASHBOARD_URL, clearLocalCredentials, friendlyHostedError, SESSION_TWIN_NAMES, runSessionCreate, runSessionList, runSessionStop, resolveCredentials, runTaskHosted, discoverRunSet, VERDICT_ARTIFACT_VERSION, persistCredentialsAfterLogin, DEFAULT_DOCS_SITE_ORIGIN, resolveSeams, resolveCachedAgentId, readLinkCache, postAgentResolver, writeLinkCache, ensurePomeGitignored } from '../../chunk-3NQUFSIX.js';
+import { DEFAULT_CONTROL_PLANE_URL, DEFAULT_DASHBOARD_URL, clearLocalCredentials, friendlyHostedError, SESSION_TWIN_NAMES, runSessionCreate, runSessionList, runSessionStop, resolveCredentials, runTaskHosted, discoverRunSet, VERDICT_ARTIFACT_VERSION, persistCredentialsAfterLogin, DEFAULT_DOCS_SITE_ORIGIN, resolveSeams, resolveCachedAgentId, readLinkCache, postAgentResolver, writeLinkCache, ensurePomeGitignored } from '../../chunk-G7252TQ3.js';
 import { assetPath, runTask, resolvePackageRoot, DEMO_TASK_NAME, demoTaskPath } from '../../chunk-AEFXTGW3.js';
 import { parseTaskFile, scoreStatus, runScoreLine, narratorReadingLines, readLatestRun, readMetaSummary, outcomeOf, readConfigTwins, scoreCountsSummary, criterionRowLine, readCodeCriteria, createHostedClient, toTwinHttpEvent, redactJsonl, scoreFromFinalizeResponse, parseGitHubSeedState, uploadRunBlobs, isPreSatisfied } from '../../chunk-BS47AE55.js';
 import '../../chunk-VPXIGCZW.js';
@@ -4575,7 +4575,7 @@ function firstSentence(description) {
 function resolveExampleRef(env = process.env) {
   const override = env.POME_EXAMPLE_REF?.trim();
   if (override) return override;
-  const baked = "ad57e84671dd2e6a2fb0dc3b66a78d97b95ed1c7".trim() ;
+  const baked = "18ff79b115d837fa3081e5925664bd24f5843c8c".trim() ;
   return FULL_SHA.test(baked) ? baked : "main";
 }
 function rawUrlFor(example, file, ref) {
@@ -4929,7 +4929,7 @@ var DEFAULT_AGENT_COMMAND = `node ${DEFAULT_AGENT_FILE}`;
 var MANIFEST_SCHEMA_URL = "https://pome.sh/schemas/v1/pome.json";
 var MAX_UNREADABLE_PATHS_SHOWN = 5;
 function readPackageVersion() {
-  if ("0.39.1".length > 0) return "0.39.1";
+  if ("0.40.0".length > 0) return "0.40.0";
   try {
     const here = dirname(fileURLToPath(import.meta.url));
     const candidates = [
@@ -4947,8 +4947,16 @@ function readPackageVersion() {
   }
   return "0.0.0";
 }
+function globals(cmd) {
+  return cmd.optsWithGlobals();
+}
 function createProgram() {
   const program = new Command();
+  program.configureHelp({ showGlobalOptions: true }).option(
+    "--api-url <url>",
+    "Control-plane base URL.",
+    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
+  ).option("--artifacts-dir <dir>", "Directory for run artifacts", "runs");
   program.name("pome").description(
     "Test AI agents against digital twins of real SaaS APIs. Runs are recorded to app.pome.sh. Start with `pome demo`."
   ).version(PACKAGE_VERSION).showHelpAfterError("(add --help for usage)");
@@ -5060,10 +5068,6 @@ function createProgram() {
     console.error(postInitMessage);
   });
   program.command("login").summary("Sign in and save an API key").description("Sign in with Clerk and store a hosted team API key (macOS Keychain or ~/.pome/credentials.json)").option(
-    "--api-url <url>",
-    "Control-plane base URL.",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
-  ).option(
     "--dashboard-url <url>",
     "App URL for Clerk sign-in (must serve /cli/login).",
     process.env.POME_DASHBOARD_URL ?? DEFAULT_DASHBOARD_URL
@@ -5072,8 +5076,8 @@ function createProgram() {
     "Label for the API key minted by this login.",
     "pome login"
   ).action(
-    async (options) => {
-      await loginWithClerk(options);
+    async (options, cmd) => {
+      await loginWithClerk({ ...options, apiUrl: globals(cmd).apiUrl });
     }
   );
   program.command("logout").summary("Delete the saved API key").description("Remove locally stored hosted credentials (Keychain entry and/or ~/.pome/credentials.json)").action(async () => {
@@ -5130,17 +5134,13 @@ function createProgram() {
     "One per declared parameter. Repeat the flag.",
     (value, previous = []) => [...previous, value],
     []
-  ).option(
-    "--api-url <url>",
-    "Control-plane URL",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).description(
     "Add one [code] criterion by picking a declared check \u2014 pome writes the English"
-  ).action(async (file, opts) => {
+  ).action(async (file, opts, cmd) => {
     await runChecksAddCommand(file, {
       check: opts.check,
       arg: opts.arg,
-      apiBaseUrl: opts.apiUrl
+      apiBaseUrl: globals(cmd).apiUrl
     });
   });
   checks.command("lint").argument("<file...>", "Task markdown file(s) \u2014 shell globs work: tasks/*.md").description("Report [code] criteria that bind no declared check, so are never graded").action(async (files) => {
@@ -5156,10 +5156,6 @@ function createProgram() {
     "Register a cloud entity (agent, ...) and link this project to it"
   );
   register.command("agent").argument("<name>", 'Human-readable agent name (e.g. "triage-bot")').option(
-    "--api-url <url>",
-    "Control-plane URL",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
-  ).option(
     "--force",
     "Re-resolve the agent even when .pome/link.json already links one",
     false
@@ -5169,10 +5165,10 @@ function createProgram() {
   ).description(
     "Create a cloud agent under the current team and write agent.slug to pome.json"
   ).action(
-    async (name, opts) => {
+    async (name, opts, cmd) => {
       try {
         await runRegisterAgent({
-          apiBaseUrl: opts.apiUrl,
+          apiBaseUrl: globals(cmd).apiUrl,
           dashboardBaseUrl: process.env.POME_DASHBOARD_URL ?? DEFAULT_DASHBOARD_URL,
           name,
           force: opts.force,
@@ -5206,18 +5202,14 @@ function createProgram() {
   ).option(
     "--seed <path>",
     "Start the sandbox from a JSON or YAML seed file instead of each twin's default. A seed REPLACES the default; it does not merge. Same file `pome twin start --seed` takes; write one with `pome twin seed <name>`."
-  ).option(
-    "--api-url <url>",
-    "Control-plane URL",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).option(
     "--secrets-file <path>",
     "Write shell exports containing session secrets to a local file with mode 0600"
   ).option("--json", "Print the sandbox as JSON instead of the human summary.", false).action(
-    async (opts) => {
+    async (opts, cmd) => {
       try {
         await runSessionCreate({
-          apiBaseUrl: opts.apiUrl,
+          apiBaseUrl: globals(cmd).apiUrl,
           twins: opts.twin ?? [],
           json: opts.json,
           secretsFile: opts.secretsFile,
@@ -5229,16 +5221,12 @@ function createProgram() {
       }
     }
   );
-  session.command("list").description("List hosted sandboxes (defaults to --state running, like the dashboard)").option(
-    "--api-url <url>",
-    "Control-plane URL",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
-  ).option("--limit <n>", "Max rows", "20").option(
+  session.command("list").description("List hosted sandboxes (defaults to --state running, like the dashboard)").option("--limit <n>", "Max rows", "20").option(
     "--state <state>",
     "Filter by sandbox state: running, ready, done, expired, or all. `running` also matches the server-side `ready` state, the way the dashboard shows them in one column.",
     "running"
   ).option("--json", "Print the sandboxes as JSON.", false).action(
-    async (opts) => {
+    async (opts, cmd) => {
       const validStates = [
         "running",
         "ready",
@@ -5255,7 +5243,7 @@ function createProgram() {
       }
       try {
         await runSessionList({
-          apiBaseUrl: opts.apiUrl,
+          apiBaseUrl: globals(cmd).apiUrl,
           limit: Number.parseInt(opts.limit, 10) || 20,
           state: opts.state,
           json: opts.json
@@ -5267,18 +5255,14 @@ function createProgram() {
     }
   );
   session.command("stop").description("Stop a hosted sandbox").argument("<session-id>", "Sandbox id (ses_\u2026)").option(
-    "--api-url <url>",
-    "Control-plane URL",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
-  ).option(
     "--discard",
     "Confirm destroying a session whose run has not been graded",
     false
   ).action(
-    async (sessionId, opts) => {
+    async (sessionId, opts, cmd) => {
       try {
         await runSessionStop({
-          apiBaseUrl: opts.apiUrl,
+          apiBaseUrl: globals(cmd).apiUrl,
           sessionId,
           discard: opts.discard === true
         });
@@ -5295,10 +5279,6 @@ function createProgram() {
   ).option("--agent <command>", "Command that starts your agent.").option(
     "-n, --trials <count>",
     "Number of trials to run as one group, 1 to 20. Hosted only; defaults to the task's `runs` field."
-  ).option("--artifacts-dir <dir>", "Directory for run artifacts", "runs").option(
-    "--api-url <url>",
-    "Control-plane base URL.",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).option("--agent-model <name>", "Informational; model name recorded on the run.", "unknown").option(
     "--agent-version <version>",
     "Override the manifest's agent.version."
@@ -5311,7 +5291,8 @@ function createProgram() {
   ).description(
     "Run a task, or every task in a directory, and print the score. With no path, runs the demo task Pome copies into your project on first use. Refuses to start when `pome doctor` fails, and there is no --force. See `pome docs cli-run` for trial groups and exit codes."
   ).action(
-    async (target, options) => {
+    async (target, options, cmd) => {
+      const { apiUrl, artifactsDir } = globals(cmd);
       let trialsFlag;
       if (options.trials !== void 0) {
         try {
@@ -5400,7 +5381,7 @@ function createProgram() {
         }
       }
       try {
-        hostedCreds = useLocal ? null : await resolveCredentials({ apiBaseUrl: options.apiUrl });
+        hostedCreds = useLocal ? null : await resolveCredentials({ apiBaseUrl: apiUrl });
       } catch (err) {
         const code = exitCodeFor(err);
         console.error(err instanceof Error ? err.message : String(err));
@@ -5424,7 +5405,7 @@ function createProgram() {
               taskForRuns.config.runs
             );
             if (k > 1) {
-              const { runTrialGroup } = await import('../../runTrialGroup-YACZ2WTV.js');
+              const { runTrialGroup } = await import('../../runTrialGroup-I6FCP2L5.js');
               const fileForRerun = relative(process.cwd(), file);
               const rerunCommand = defaultTask ? options.trials !== void 0 ? `pome run -n ${k}` : "pome run" : `pome run ${fileForRerun && !fileForRerun.startsWith("..") ? fileForRerun : file} -n ${k}`;
               const groupResult = await runTrialGroup({
@@ -5432,7 +5413,7 @@ function createProgram() {
                 agentCommand,
                 agentCommandSource: options.agent ? "--agent" : configCommand ? "pome.json" : "built-in default",
                 trials: k,
-                artifactsDir: options.artifactsDir,
+                artifactsDir,
                 hosted: {
                   baseUrl: hostedCreds.apiBaseUrl,
                   apiKey: hostedCreds.apiKey
@@ -5450,7 +5431,7 @@ function createProgram() {
             const result = await runTaskHosted({
               taskPath: file,
               agentCommand,
-              artifactsDir: options.artifactsDir,
+              artifactsDir,
               hosted: { baseUrl: hostedCreds.apiBaseUrl, apiKey: hostedCreds.apiKey },
               agentModel: options.agentModel,
               agentVersion: options.agentVersion
@@ -5478,7 +5459,7 @@ function createProgram() {
           const result = await runTask({
             taskPath: file,
             agentCommand,
-            artifactsDir: options.artifactsDir,
+            artifactsDir,
             // Commander negates --no-* flags: `--no-capture` → `capture: false`.
             noCapture: options.capture === false
           });
@@ -5505,16 +5486,12 @@ function createProgram() {
   );
   program.command("demo").summary("Run a sample task, no account needed").description(
     "Zero-auth first-run demo: boots a local GitHub twin, runs the bundled demo agent for 5 isolated trials (model calls via pome's anonymous demo gateway), and prints per-trial verdicts evaluated in Pome cloud. No signup, no API keys; ends with a no-login preview link."
-  ).option(
-    "--api-url <url>",
-    "Control-plane base URL.",
-    process.env.POME_API_BASE ?? process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).option(
     "--trials <n>",
     "Number of trials to run, 1 to 10",
     "5"
-  ).option("--artifacts-dir <dir>", "Directory for run artifacts", "runs").action(
-    async (opts) => {
+  ).action(
+    async (opts, cmd) => {
       const trials = Number.parseInt(opts.trials, 10);
       if (!Number.isInteger(trials) || trials < 1 || trials > 10) {
         console.error(`Invalid --trials "${opts.trials}" (expected 1-10).`);
@@ -5523,10 +5500,10 @@ function createProgram() {
       }
       const { runDemo } = await import('../../runDemo-B2SXM4OC.js');
       const result = await runDemo({
-        apiBase: opts.apiUrl.replace(/\/$/, ""),
+        apiBase: globals(cmd).apiUrl.replace(/\/$/, ""),
         dashboardBase: process.env.POME_DASHBOARD_URL ?? DEFAULT_DASHBOARD_URL,
         trials,
-        artifactsDir: opts.artifactsDir
+        artifactsDir: globals(cmd).artifactsDir
       });
       process.exitCode = result.exitCode;
     }
@@ -5548,29 +5525,21 @@ function createProgram() {
   program.command("eval").summary("Score a trace recorded earlier").argument(
     "[run-dir]",
     "Existing run directory (runs/<task>/<run-id>). Omit to use <artifacts-dir>/latest.json."
-  ).option(
-    "--artifacts-dir <dir>",
-    "Directory whose latest.json picks the run when no run dir is given.",
-    "runs"
   ).option(
     "--agent <slug>",
     "Agent identity for the eval session. Defaults to agent.slug from pome.json."
   ).option(
     "--task <name>",
     "Task name recorded on the eval session. Defaults to meta.json's `scenario` slug (a legacy key name; then title)."
-  ).option(
-    "--api-url <url>",
-    "Control-plane base URL.",
-    process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).description(
     "Upload an existing raw trace directory to Pome cloud for evaluation and print the authoritative score (capture/eval split \u2014 no local scoring; requires a control plane with POST /v1/eval-sessions)"
   ).action(
-    async (runDir, opts) => {
-      await runEvalCommand(runDir, opts);
+    async (runDir, opts, cmd) => {
+      await runEvalCommand(runDir, { ...opts, ...globals(cmd) });
     }
   );
-  program.command("inspect").argument("<run>", "Run id, run directory, or latest").option("--artifacts-dir <dir>", "Directory for run artifacts", "runs").description("Print a human-readable run report").action(async (run, options) => {
-    const latest = run === "latest" ? await readLatestRun(options.artifactsDir) : void 0;
+  program.command("inspect").argument("<run>", "Run id, run directory, or latest").description("Print a human-readable run report").action(async (run, _options, cmd) => {
+    const latest = run === "latest" ? await readLatestRun(globals(cmd).artifactsDir) : void 0;
     const runDir = latest?.run_dir ?? resolve(run);
     const eventsResult = await readEventsJsonl(runDir);
     const meta = await readMetaSummary(runDir);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pome-sh/cli",
-  "version": "0.39.1",
+  "version": "0.40.0",
   "description": "Test AI agents against digital twins of real SaaS APIs. Records tool-call traces and scores them on pome.sh.",
   "keywords": [
     "ai",