@pome-sh/cli 0.33.1 → 0.33.3

package/dist/build-info.json

@@ -1,6 +1,6 @@
 {
   "package": "pome-sh",
-  "version": "0.33.1",
-  "git_sha": "15d61b80f3fb2ee0df893ca81c0c9e4d0afd0076",
-  "build_time": "2026-08-28T11:26:59.938Z"
+  "version": "0.33.3",
+  "git_sha": "36fe4446b6e9ba8fa1c839cabeeae68e4b7fd929",
+  "build_time": "2026-08-28T12:24:00.501Z"
 }

package/dist/src/cli/main.js

@@ -4622,7 +4622,7 @@ 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 ("0.33.1".length > 0) return "0.33.1";
+  if ("0.33.3".length > 0) return "0.33.3";
   try {
     const here = dirname(fileURLToPath(import.meta.url));
     const candidates = [
@@ -4888,7 +4888,7 @@ function createProgram() {
     "Write shell exports containing session secrets to a local file with mode 0600"
   ).option(
     "--format <fmt>",
-    "text (default) | json | env. env requires --secrets-file and is not printed.",
+    "Output format: text, json, or env. env writes the exports to --secrets-file instead of printing them.",
     "text"
   ).action(
     async (opts) => {
@@ -4918,9 +4918,9 @@ function createProgram() {
     process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).option("--limit <n>", "Max rows", "20").option(
     "--state <state>",
-    "Filter by session state: running (default) | ready | done | expired | all. `running` matches both server-side `ready` and `running` (dashboard collapses both into one column).",
+    "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("--format <fmt>", "text | json", "text").action(
+  ).option("--format <fmt>", "Output format: text or json", "text").action(
     async (opts) => {
       const validStates = [
         "running",
@@ -4974,25 +4974,25 @@ function createProgram() {
   );
   program.command("run").summary("Run a task and print the score").argument(
     "[path]",
-    'Task markdown file or directory. Omit to run the demo task ("that was ours, run yours"): tasks/first-run-demo.md is dropped into your project on first use with runs: 5 pinned.'
-  ).option("--agent <command>", "Agent command to run").option(
+    "Task markdown file or directory. Omit to run the demo task Pome copies into your project."
+  ).option("--agent <command>", "Command that starts your agent.").option(
     "-n, --trials <count>",
-    "Run <count> isolated trials of the task as ONE trial group (integer 1-20; hosted only). Default is the task config's `runs` field (capped at 20). k>1 mints sessions with a shared group id up to your plan's concurrent-twin quota (remaining trials reuse slots as earlier trials finish), runs trials at that concurrency, prints the per-trial verdict table (numeric cloud-judge scores), and exits 0 iff at least one trial completed and every completed trial passed (1: a completed trial failed; 2: nothing completed). k=1 keeps today's single-run behavior exactly."
+    "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; recorded on the cloud run.", "unknown").option(
+  ).option("--agent-model <name>", "Informational; model name recorded on the run.", "unknown").option(
     "--agent-version <version>",
-    "Override the manifest's agent.version for this run (stamped on the session/run)."
+    "Override the manifest's agent.version."
   ).option(
     "--no-capture",
-    "Self-host only: skip spawning the capture-server child and don't inject HTTP_PROXY/HTTPS_PROXY into the agent. Used by the CI overhead gate to baseline proxy-on-vs-off latency. No-op on hosted runs."
+    "Self-host only: skip the capture proxy, so model calls are not recorded. Ignored on hosted runs."
   ).option(
     "--local",
-    "Self-host: run against an in-process twin and CAPTURE a raw trace only (an audit log \u2014 no score, no verdict, no judge). A verdict comes from the cloud: run `pome eval <run-dir>` on the captured trace, or `pome login` and run against Pome cloud."
+    "Self-host: record a trace against an in-process twin, with no score. Score it with `pome eval`."
   ).description(
-    'Run one or more Pome tasks. With no path, runs the demo task (tasks/first-run-demo.md, copied into your project on first use \u2014 "that was ours, run yours"). Refuses to start if the doctor wiring checks fail (see `pome doctor`); there is no --force.'
+    "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) => {
       let trialsFlag;
@@ -5191,7 +5191,7 @@ function createProgram() {
     process.env.POME_API_BASE ?? process.env.POME_API_URL ?? DEFAULT_CONTROL_PLANE_URL
   ).option(
     "--trials <n>",
-    "Number of isolated trials (default 5 per the packaged demo).",
+    "Number of trials to run, 1 to 10",
     "5"
   ).option("--artifacts-dir <dir>", "Directory for run artifacts", "runs").action(
     async (opts) => {
@@ -5408,7 +5408,9 @@ function createProgram() {
     const { runTwinSeedCommand } = await import('../../twinSeed-2MRVMNIT.js');
     await runTwinSeedCommand(names, options);
   });
-  twin.command("reset").argument("[name]", "Twin name (default: github)", "github").description("Reset standalone twin state").action(async (name) => {
+  twin.command("reset").argument("[name]", "Twin name", "github").summary("Reset a twin's saved state").description(
+    "Delete the twin's database under `.pome/` or `.pome-data/`, and `.pome/twin-status.json`, the one file `pome twin status` reads for every twin. A twin holding its state in memory has nothing to delete."
+  ).action(async (name) => {
     if (!isTwinName(name)) {
       throw new Error(`Unknown twin '${name}'. Supported: ${TWIN_NAME_LIST.join(", ")}.`);
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pome-sh/cli",
-  "version": "0.33.1",
+  "version": "0.33.3",
   "description": "Test AI agents against digital twins of real SaaS APIs. Records tool-call traces and scores them on pome.sh.",
   "keywords": [
     "ai",