@wrongstack/cli 0.319.1 → 1.0.1

package/dist/{chunk-XPX432BZ.js → chunk-DR2OXU46.js}

@@ -107,47 +107,58 @@ async function loadConfigProviders(configPath, vault, opts) {
   const decrypted = decryptConfigSecrets(parsed, vault);
   return decrypted.providers ?? {};
 }
+var configMutationChain = Promise.resolve();
+function queueConfigMutation(mutate) {
+  const run = configMutationChain.then(mutate, mutate);
+  configMutationChain = run.then(
+    () => void 0,
+    () => void 0
+  );
+  return run;
+}
 async function mutateConfigProviders(configPath, vault, mutator, profileConfigPath) {
-  const targetPath = profileConfigPath ?? configPath;
-  let raw;
-  let fileExists = true;
-  try {
-    raw = await fs.readFile(targetPath, "utf8");
-  } catch (err) {
-    if (err.code !== "ENOENT") {
-      throw new FsError({
-        message: `Refusing to mutate ${configPath}: ${err.message}`,
-        code: "FS_READ_FAILED",
-        path: targetPath,
-        context: { operation: "mutateConfigProviders", phase: "read" },
-        cause: err
-      });
+  await queueConfigMutation(async () => {
+    const targetPath = profileConfigPath ?? configPath;
+    let raw;
+    let fileExists = true;
+    try {
+      raw = await fs.readFile(targetPath, "utf8");
+    } catch (err) {
+      if (err.code !== "ENOENT") {
+        throw new FsError({
+          message: `Refusing to mutate ${configPath}: ${err.message}`,
+          code: "FS_READ_FAILED",
+          path: targetPath,
+          context: { operation: "mutateConfigProviders", phase: "read" },
+          cause: err
+        });
+      }
+      fileExists = false;
+      raw = "{}";
     }
-    fileExists = false;
-    raw = "{}";
-  }
-  let parsed;
-  try {
-    parsed = JSON.parse(raw);
-  } catch (err) {
-    if (fileExists) {
-      throw new FsError({
-        message: `Refusing to overwrite corrupt config at ${targetPath} (${err.message}). Fix or move the file aside before retrying.`,
-        code: "FS_READ_FAILED",
-        path: targetPath,
-        context: { operation: "mutateConfigProviders", phase: "parse" },
-        cause: err
-      });
+    let parsed;
+    try {
+      parsed = JSON.parse(raw);
+    } catch (err) {
+      if (fileExists) {
+        throw new FsError({
+          message: `Refusing to overwrite corrupt config at ${targetPath} (${err.message}). Fix or move the file aside before retrying.`,
+          code: "FS_READ_FAILED",
+          path: targetPath,
+          context: { operation: "mutateConfigProviders", phase: "parse" },
+          cause: err
+        });
+      }
+      parsed = {};
     }
-    parsed = {};
-  }
-  const decrypted = decryptConfigSecrets(parsed, vault);
-  const providers = decrypted.providers ?? {};
-  mutator(providers);
-  decrypted.providers = providers;
-  clearStaleProviderDefaults(decrypted);
-  const encrypted = encryptConfigSecrets(decrypted, vault);
-  await atomicWrite(targetPath, JSON.stringify(encrypted, null, 2), { mode: 384 });
+    const decrypted = decryptConfigSecrets(parsed, vault);
+    const providers = decrypted.providers ?? {};
+    mutator(providers, decrypted);
+    decrypted.providers = providers;
+    clearStaleProviderDefaults(decrypted);
+    const encrypted = encryptConfigSecrets(decrypted, vault);
+    await atomicWrite(targetPath, JSON.stringify(encrypted, null, 2), { mode: 384 });
+  });
 }
 
 export {
@@ -160,4 +171,4 @@ export {
   mutateConfigProviders,
   nowIso
 };
-//# sourceMappingURL=chunk-XPX432BZ.js.map
+//# sourceMappingURL=chunk-DR2OXU46.js.map

package/dist/{chunk-GGGUZSAU.js → chunk-E6RFXHWD.js}

@@ -22,7 +22,6 @@ var BOOLEAN_FLAGS = /* @__PURE__ */ new Set([
   "no-recovery",
   "recover",
   "output-json",
-  "prompt",
   "metrics",
   "webui",
   "webui-session-child",
@@ -36,8 +35,6 @@ var BOOLEAN_FLAGS = /* @__PURE__ */ new Set([
   "no-models-refresh",
   // (removed: 'director' and 'no-director' — Director Mode is always on)
   "no-autonomy",
-  "no-hints",
-  "hints",
   "no-menu",
   "no-hooks",
   "skip",
@@ -174,20 +171,28 @@ function parseAuthFlags(args) {
   const out = { positional: [] };
   for (let i = 0; i < args.length; i++) {
     const a = args[i];
-    if (a === "--label") {
-      const v = args[++i];
+    if (!a) continue;
+    let key = a;
+    let inlineVal;
+    const eq = a.indexOf("=");
+    if (a.startsWith("--") && eq !== -1) {
+      key = a.slice(0, eq);
+      inlineVal = a.slice(eq + 1);
+    }
+    if (key === "--label") {
+      const v = inlineVal !== void 0 ? inlineVal : args[++i];
       if (v) out.label = v;
-    } else if (a === "--family") {
-      const v = args[++i];
+    } else if (key === "--family") {
+      const v = inlineVal !== void 0 ? inlineVal : args[++i];
       if (v) out.family = v;
-    } else if (a === "--base-url") {
-      const v = args[++i];
+    } else if (key === "--base-url") {
+      const v = inlineVal !== void 0 ? inlineVal : args[++i];
       if (v) out.baseUrl = v;
-    } else if (a === "--env") {
-      const v = args[++i];
+    } else if (key === "--env") {
+      const v = inlineVal !== void 0 ? inlineVal : args[++i];
       if (v)
         out.envVars = v.split(",").map((s) => s.trim()).filter(Boolean);
-    } else if (a && !a.startsWith("--")) {
+    } else if (!a.startsWith("--")) {
       out.positional.push(a);
     }
   }
@@ -232,4 +237,4 @@ export {
   parseAuthFlags,
   parseSpawnFlags
 };
-//# sourceMappingURL=chunk-GGGUZSAU.js.map
+//# sourceMappingURL=chunk-E6RFXHWD.js.map

package/dist/{chunk-TYO2OVAD.js → chunk-EWKESL4L.js}

@@ -11,9 +11,13 @@ function patchConfig(base, patch) {
 function fmtDuration(ms) {
   if (ms < 1e3) return `${ms}ms`;
   const s = ms / 1e3;
-  if (s < 60) return `${s.toFixed(s < 10 ? 1 : 0)}s`;
-  const m = Math.floor(s / 60);
-  const remSec = Math.round(s - m * 60);
+  if (s < 59.5) return `${s.toFixed(s < 10 ? 1 : 0)}s`;
+  let m = Math.floor(s / 60);
+  let remSec = Math.round(s - m * 60);
+  if (remSec === 60) {
+    remSec = 0;
+    m++;
+  }
   if (m < 60) return remSec === 0 ? `${m}m` : `${m}m${remSec}s`;
   const h = Math.floor(m / 60);
   const remMin = m - h * 60;
@@ -47,4 +51,4 @@ export {
   patchConfig,
   fmtTaskResultLine
 };
-//# sourceMappingURL=chunk-TYO2OVAD.js.map
+//# sourceMappingURL=chunk-EWKESL4L.js.map

package/dist/{chunk-QAC263FZ.js → chunk-FPIWTJ2G.js}

@@ -55,7 +55,7 @@ var helpCmd = async (_args, deps) => {
     "  wstack mailbox serve [opts]  Start external-agent mailbox HTTP bridge",
     "  wstack permissions explain   Explain tool permission policy decisions",
     "  wstack modeldiag [test]      Model benchmarks and capability diagnostics",
-    "  wstack bench [run|list]      Agentic benchmarks against standard suites",
+    "  wstack bench [run|compare]   Agentic benchmarks against standard suites",
     "  wstack acp [serve]           Agent Client Protocol (ACP) server",
     "  wstack update [--check-only] Self-update the CLI",
     "  wstack diag                  Full diagnostics",
@@ -81,7 +81,6 @@ var helpCmd = async (_args, deps) => {
     "                               Serve the independent minimal chat UI",
     "  --full-auto                   SimpleUI: runtime-only YOLO + Director + autonomy override",
     '  --eternal "<mission>"        Start an eternal-autonomy loop',
-    "  --no-hints                   Hide launch hints",
     "  --skip-index                 Skip codebase indexing on startup",
     "  --chimera-auto-fix off|ask|auto",
     "                               How to handle Chimera review findings (default: config value)"
@@ -94,4 +93,4 @@ export {
   versionCmd,
   helpCmd
 };
-//# sourceMappingURL=chunk-QAC263FZ.js.map
+//# sourceMappingURL=chunk-FPIWTJ2G.js.map

package/dist/{chunk-7YLN7YUA.js → chunk-IMWXXF4J.js}

@@ -20,7 +20,34 @@ function createHqCommandDispatcher(controller) {
     }
   };
 }
+function resolveCommandSessionId(controller, payload) {
+  const named = payload["sessionId"];
+  if (typeof named === "string" && named.length > 0) return named;
+  return controller.sessionId?.();
+}
+function rejectUnknownSession(controller, commandId, sessionId) {
+  const named = sessionId;
+  if (named === void 0) return null;
+  const owns = controller.ownsSession?.(named) ?? (controller.sessionId === void 0 ? true : controller.sessionId() === named);
+  if (owns) return null;
+  return {
+    commandId,
+    status: "rejected",
+    message: `session ${named} is not active on this client`
+  };
+}
+function leaderSessionAffinity(to, sessionId) {
+  if (to.trim().toLowerCase() !== "leader") return {};
+  if (typeof sessionId !== "string" || sessionId.length === 0) return {};
+  return { sessionAffinity: { sessionId } };
+}
 async function dispatch(controller, commandId, type, payload) {
+  const sessionId = resolveCommandSessionId(controller, payload);
+  const namedSession = typeof payload["sessionId"] === "string" ? payload["sessionId"] : void 0;
+  if (type !== "broadcast") {
+    const rejected = rejectUnknownSession(controller, commandId, namedSession);
+    if (rejected !== null) return rejected;
+  }
   switch (type) {
     case "steer": {
       const mailbox = controller.steerMailbox;
@@ -32,7 +59,15 @@ async function dispatch(controller, commandId, type, payload) {
       const body = typeof payload["body"] === "string" ? payload["body"] : "";
       const priority = payload["priority"] === "high" ? "high" : payload["priority"] === "low" ? "low" : "normal";
       const from = `hq@${controller.sessionTag()}`;
-      await mailbox.send({ from, to, type: "steer", subject, body, priority });
+      await mailbox.send({
+        from,
+        to,
+        type: "steer",
+        subject,
+        body,
+        priority,
+        ...leaderSessionAffinity(to, sessionId)
+      });
       return { commandId, status: "accepted", message: `steered ${to}` };
     }
     case "btw": {
@@ -45,7 +80,15 @@ async function dispatch(controller, commandId, type, payload) {
       const body = typeof payload["body"] === "string" ? payload["body"] : "";
       const priority = payload["priority"] === "high" ? "high" : payload["priority"] === "low" ? "low" : "normal";
       const from = `hq@${controller.sessionTag()}`;
-      await mailbox.send({ from, to, type: "btw", subject, body, priority });
+      await mailbox.send({
+        from,
+        to,
+        type: "btw",
+        subject,
+        body,
+        priority,
+        ...leaderSessionAffinity(to, sessionId)
+      });
       return { commandId, status: "accepted", message: `noted ${to}` };
     }
     case "queue": {
@@ -58,7 +101,15 @@ async function dispatch(controller, commandId, type, payload) {
       const body = typeof payload["body"] === "string" ? payload["body"] : "";
       const priority = payload["priority"] === "high" ? "high" : payload["priority"] === "low" ? "low" : "normal";
       const from = `hq@${controller.sessionTag()}`;
-      await mailbox.send({ from, to, type: "note", subject, body, priority });
+      await mailbox.send({
+        from,
+        to,
+        type: "note",
+        subject,
+        body,
+        priority,
+        ...leaderSessionAffinity(to, sessionId)
+      });
       return { commandId, status: "accepted", message: `queued for ${to}` };
     }
     case "broadcast": {
@@ -76,7 +127,7 @@ async function dispatch(controller, commandId, type, payload) {
     case "abort": {
       const target = typeof payload["target"] === "string" ? payload["target"] : "leader";
       if (target === "leader") {
-        const aborted = controller.interruptLeader();
+        const aborted = controller.interruptLeader(sessionId);
         return {
           commandId,
           status: aborted ? "completed" : "accepted",
@@ -84,14 +135,14 @@ async function dispatch(controller, commandId, type, payload) {
         };
       }
       if (target === "fleet") {
-        const killed = await controller.killFleet?.() ?? 0;
+        const killed = await controller.killFleet?.(sessionId) ?? 0;
         return {
           commandId,
           status: "completed",
           message: `fleet stopped (${killed} agents)`
         };
       }
-      const terminated = await controller.terminateAgent?.(target) ?? false;
+      const terminated = await controller.terminateAgent?.(target, sessionId) ?? false;
       return {
         commandId,
         status: terminated ? "completed" : "rejected",
@@ -105,7 +156,7 @@ async function dispatch(controller, commandId, type, payload) {
       if (controller.spawnAgent === void 0) {
         return { commandId, status: "rejected", message: "no director active" };
       }
-      const subagentId = await controller.spawnAgent(role, task, maxIterations);
+      const subagentId = await controller.spawnAgent(role, task, maxIterations, sessionId);
       return { commandId, status: "completed", message: `spawned ${subagentId}` };
     }
     case "run-command": {
@@ -130,7 +181,8 @@ async function dispatch(controller, commandId, type, payload) {
 \`\`\`
 ${command}
 \`\`\``,
-          priority: "high"
+          priority: "high",
+          ...leaderSessionAffinity("leader", sessionId)
         });
         return { commandId, status: "accepted", message: "run-command routed to leader as steer" };
       }
@@ -144,4 +196,4 @@ ${command}
 export {
   createHqCommandDispatcher
 };
-//# sourceMappingURL=chunk-7YLN7YUA.js.map
+//# sourceMappingURL=chunk-IMWXXF4J.js.map

package/dist/{chunk-LIWVIEZM.js → chunk-KHWDTNBZ.js}

@@ -1,10 +1,10 @@
 import {
   helpCmd,
   versionCmd
-} from "./chunk-QAC263FZ.js";
+} from "./chunk-FPIWTJ2G.js";
 import {
   parseArgs
-} from "./chunk-GGGUZSAU.js";
+} from "./chunk-E6RFXHWD.js";
 
 // src/boot/short-circuit-flags.ts
 async function handleHelpVersionShortCircuit(argv) {
@@ -22,11 +22,11 @@ async function handleHelpVersionShortCircuit(argv) {
   }
   if (earlyFlags["help"] === true) {
     if (positional.length > 0) {
-      const { subcommandNames } = await import("./subcommands-KE6MEQQP.js");
+      const { subcommandNames } = await import("./subcommands-WJLBHVKC.js");
       if (subcommandNames.includes(positional[0])) {
         const first = positional[0];
         const deep = positional[1];
-        const { renderDeepHelp, renderFocusedHelp } = await import("./per-subcommand-help-6VCIKDUW.js");
+        const { renderDeepHelp, renderFocusedHelp } = await import("./per-subcommand-help-BWBHYVJO.js");
         if (deep && renderDeepHelp(`${first}:${deep}`, stubRenderer)) {
           return 0;
         }
@@ -43,4 +43,4 @@ async function handleHelpVersionShortCircuit(argv) {
 export {
   handleHelpVersionShortCircuit
 };
-//# sourceMappingURL=chunk-LIWVIEZM.js.map
+//# sourceMappingURL=chunk-KHWDTNBZ.js.map

package/dist/{chunk-IMT6PNV5.js → chunk-MNQQ4ZNF.js}

@@ -11,7 +11,7 @@ import {
   openBrowser,
   runCodexOAuthLogin,
   startLoopbackServer
-} from "./chunk-E4I2TF5S.js";
+} from "./chunk-TNK6237M.js";
 import {
   activeLabel,
   loadConfigProviders,
@@ -20,7 +20,7 @@ import {
   normalizeKeys,
   nowIso,
   writeKeysBack
-} from "./chunk-XPX432BZ.js";
+} from "./chunk-DR2OXU46.js";
 
 // src/auth-menu/local.ts
 import { DefaultSecretScrubber as DefaultSecretScrubber2 } from "@wrongstack/core/security";
@@ -1444,11 +1444,11 @@ function renderLiveProviderList(query, filtered, selectedIdx, savedSet) {
 }
 async function saveToGlobalConfig(configPath, provider, model, homeFn = () => process.env.HOME ?? os.homedir()) {
   try {
-    const { atomicWrite: atomicWrite2 } = await import("@wrongstack/core/utils");
-    const fs2 = await import("node:fs/promises");
+    const { atomicWrite } = await import("@wrongstack/core/utils");
+    const fs = await import("node:fs/promises");
     let existing = {};
     try {
-      const raw = await fs2.readFile(configPath, "utf8");
+      const raw = await fs.readFile(configPath, "utf8");
       existing = JSON.parse(raw);
     } catch {
     }
@@ -1467,7 +1467,7 @@ async function saveToGlobalConfig(configPath, provider, model, homeFn = () => pr
         })
       );
     }
-    await atomicWrite2(configPath, JSON.stringify(existing, null, 2), { mode: 384 });
+    await atomicWrite(configPath, JSON.stringify(existing, null, 2), { mode: 384 });
     try {
       await appendHistory(
         oldCfg,
@@ -1899,8 +1899,7 @@ async function resolveModelSelection(answer, models, provider, _registry, render
 }
 
 // src/auth-menu/add-provider.ts
-import * as fs from "node:fs/promises";
-import { atomicWrite, color as color9 } from "@wrongstack/core/utils";
+import { color as color9 } from "@wrongstack/core/utils";
 import { catalogProviderIdFor, PROVIDER_DEFINITIONS } from "@wrongstack/providers";
 
 // src/provider-id.ts
@@ -2148,22 +2147,26 @@ async function resolveDefaultModelId(deps, providerId, template) {
   return provider?.models?.[0]?.id;
 }
 async function adoptAsDefaultIfUnset(deps, providerId, template) {
-  const targetPath = deps.profileConfigPath;
-  let cfg = {};
-  try {
-    cfg = JSON.parse(await fs.readFile(targetPath, "utf8"));
-  } catch {
-  }
-  if (cfg["provider"] && cfg["model"]) return;
   const modelId = await resolveDefaultModelId(deps, providerId, template);
   if (!modelId) return;
-  cfg["provider"] = providerId;
-  cfg["model"] = modelId;
-  await atomicWrite(targetPath, JSON.stringify(cfg, null, 2), { mode: 384 });
-  deps.renderer.write(
-    `  ${color9.green("\u2713")} Set ${color9.bold(providerId)}/${color9.bold(modelId)} as the default.
-`
+  let adopted = false;
+  await mutateConfigProviders(
+    deps.profileConfigPath,
+    deps.vault,
+    (_providers, config) => {
+      if (config["provider"] && config["model"]) return;
+      config["provider"] = providerId;
+      config["model"] = modelId;
+      adopted = true;
+    },
+    deps.profileConfigPath
   );
+  if (adopted) {
+    deps.renderer.write(
+      `  ${color9.green("\u2713")} Set ${color9.bold(providerId)}/${color9.bold(modelId)} as the default.
+`
+    );
+  }
 }
 async function promptForLabel(deps, usedLabels) {
   const defaultLabel = suggestLabel(usedLabels);
@@ -2201,9 +2204,10 @@ export {
   saveToGlobalConfig,
   runPicker,
   nextCustomProviderId,
+  ownDefinitionsAsCatalog,
   addFromCatalog,
   addKeyForCatalogProvider,
   addCustomProvider,
   addKeyForProvider
 };
-//# sourceMappingURL=chunk-IMT6PNV5.js.map
+//# sourceMappingURL=chunk-MNQQ4ZNF.js.map

package/dist/{chunk-P2OKBDCH.js → chunk-MXEJEYT7.js}

@@ -12,18 +12,17 @@ var BENCH_RUN_FLAGS = [
   {
     name: "suite",
     flag: "--suite <id>",
-    description: "Benchmark suite id (`polyglot` or `swebench`).",
+    description: "Benchmark suite id (`core`, `smoke`, `local`, `polyglot`, or `swebench`).",
     group: "suite",
     kind: "value",
-    defaultValue: "polyglot"
+    defaultValue: "core"
   },
   {
     name: "polyglot-dir",
     flag: "--polyglot-dir <path>",
     description: "Path to the Aider polyglot dataset (required when --suite polyglot).",
     group: "suite",
-    kind: "value",
-    required: true
+    kind: "value"
   },
   {
     name: "languages",
@@ -46,15 +45,36 @@ var BENCH_RUN_FLAGS = [
     group: "suite",
     kind: "boolean"
   },
+  {
+    name: "suite-dir",
+    flag: "--suite-dir <path>",
+    description: "Directory containing bench.local.json (required when --suite local).",
+    group: "suite",
+    kind: "value"
+  },
+  {
+    name: "manifest",
+    flag: "--manifest <file>",
+    description: "Explicit local-manifest path (alternative to --suite-dir).",
+    group: "suite",
+    kind: "value"
+  },
   // -- Model matrix ---------------------------------------------------
   {
     name: "models",
     flag: "--models <config>",
-    description: "Path to the model-cells config (JSON).",
+    description: "Path to the model-cells config (JSON). Optional when --cell or a saved model is set.",
     group: "models",
     kind: "value",
     defaultValue: "bench.config.json"
   },
+  {
+    name: "cell",
+    flag: "--cell <spec>",
+    description: "Comma-separated model cells (`provider/model` or `label=provider/model`). Skips the config file.",
+    group: "models",
+    kind: "value"
+  },
   // -- Run control ----------------------------------------------------
   {
     name: "limit",
@@ -77,6 +97,21 @@ var BENCH_RUN_FLAGS = [
     description: "Override the per-cell concurrency from the model config.",
     group: "control",
     kind: "value"
+  },
+  {
+    name: "repeats",
+    flag: "--repeats <N>",
+    description: "Attempts per (task x model). >1 adds pass@k and a flakiness count so run-to-run noise is visible.",
+    group: "control",
+    kind: "value",
+    defaultValue: "1"
+  },
+  {
+    name: "keep-sandbox",
+    flag: "--keep-sandbox",
+    description: "Keep the temporary sandbox (workdirs + isolated home) on disk for debugging.",
+    group: "control",
+    kind: "boolean"
   }
 ];
 var BENCH_RUN_BOOLEAN_FLAG_NAMES = BENCH_RUN_FLAGS.filter(
@@ -89,10 +124,10 @@ var BENCH_RUN_FLAG_COLUMN_WIDTH = 28;
 function renderBenchRunHelpToString() {
   const lines = [
     color.bold("wstack bench run \u2014 execute a benchmark suite across a model matrix"),
-    color.dim("  Runs a benchmark suite (polyglot or swebench) across every model cell"),
-    color.dim("  in the config. Output is a per-run directory with `report.md`, JSON"),
-    color.dim("  artifacts, and (for swebench) per-cell predictions for the official"),
-    color.dim("  harness."),
+    color.dim("  Default suite is bundled core (6 agent-edit tasks with tests). Pass --cell"),
+    color.dim("  provider/model to start immediately. Output is a per-run directory"),
+    color.dim("  with `report.md` (leaderboard + per-task matrix), JSON artifacts,"),
+    color.dim("  and (for swebench) per-cell predictions for the official harness."),
     "",
     color.bold("Usage"),
     `  ${buildUsageLine()}`,
@@ -101,7 +136,7 @@ function renderBenchRunHelpToString() {
     ...buildFlagBlock(),
     "",
     color.dim(
-      "See also: wstack bench list (show available suites + cells); wstack bench report <dir>"
+      "See also: wstack bench list; wstack bench report <dir>; wstack bench compare <a> <b>"
     )
   ];
   return lines.join("\n") + "\n";
@@ -969,10 +1004,14 @@ var helpTable = {
   bench: {
     name: "bench",
     title: "wstack bench \u2014 run model-independent agentic benchmarks",
-    description: "Run WrongStack against the Aider polyglot or SWE-bench Verified suites with deterministic graders. Used internally to compare model quality across releases; also useful for evaluating a new model before adopting it.",
-    usage: "wstack bench [run|report|list] [...]",
+    description: "Run WrongStack against the bundled smoke suite, a local manifest, Aider polyglot, or SWE-bench Verified with deterministic graders. Default `bench run` is the bundled core edit-eval; pass --cell provider/model (or a saved model).",
+    usage: "wstack bench [run|compare|report|list] [...]",
     subcommands: [
-      { name: "run", description: "Run a benchmark suite (--suite <id> --models <config>)." },
+      { name: "run", description: "Run a suite. Default: bundled smoke + --cell or saved model." },
+      {
+        name: "compare <a> <b>",
+        description: "Diff two finished run directories (fingerprint-aware)."
+      },
       { name: "report <dir>", description: "Render the Markdown report for a prior run." },
       { name: "list", description: "List available suites and the model configs in the catalog." }
     ],
@@ -1163,4 +1202,4 @@ export {
   subcommandsWithFocusedHelp,
   renderGenericHelp
 };
-//# sourceMappingURL=chunk-P2OKBDCH.js.map
+//# sourceMappingURL=chunk-MXEJEYT7.js.map

package/dist/{chunk-PCBJSXYV.js → chunk-O4A3HHKD.js}

@@ -14,7 +14,7 @@ async function printUpdateNotice(initialUpdateInfo) {
     const ac = new AbortController();
     const timer = setTimeout(() => ac.abort(), 2e3);
     try {
-      const { checkForUpdate } = await import("./update-check-PKSMSMD6.js");
+      const { checkForUpdate } = await import("./update-check-BS3DHTSI.js");
       info = await checkForUpdate(ac.signal);
     } catch {
     } finally {
@@ -62,4 +62,4 @@ export {
   applySessionShellDefault,
   runPreflight
 };
-//# sourceMappingURL=chunk-PCBJSXYV.js.map
+//# sourceMappingURL=chunk-O4A3HHKD.js.map

package/dist/{chunk-XYKYXFEE.js → chunk-TE3Q32JP.js}

@@ -423,6 +423,7 @@ async function writeState(projectRoot, state) {
 }
 
 // src/hq-publisher.ts
+var HQ_DISCOVERY_CONNECTION_KEY = "<discover>";
 function nodeWsSocketFactory(url) {
   return new WebSocket(url);
 }
@@ -440,11 +441,13 @@ function hqExplicitlyDisabled(options) {
 }
 function resolvedHqConnectionKey(options) {
   if (options.config !== void 0) {
-    return options.config.enabled === false ? void 0 : `${options.config.url}
+    if (options.config.enabled === false) return void 0;
+    return options.config.discover === true ? HQ_DISCOVERY_CONNECTION_KEY : `${options.config.url}
 ${options.config.token ?? ""}`;
   }
   const config = resolveHqConfig({ config: options.appConfig?.hq });
-  return config === void 0 ? void 0 : `${config.url}
+  if (config === void 0) return void 0;
+  return config.discover === true ? HQ_DISCOVERY_CONNECTION_KEY : `${config.url}
 ${config.token ?? ""}`;
 }
 function startCliHqConnection(options) {
@@ -460,7 +463,10 @@ function startCliHqConnection(options) {
   const tryConnect = () => {
     const nextKey = resolvedHqConnectionKey(options);
     if (nextKey === void 0) return;
-    if (publisher !== void 0 && publisherKey === nextKey) return;
+    if (publisher !== void 0 && publisherKey === nextKey) {
+      if (nextKey === HQ_DISCOVERY_CONNECTION_KEY) publisher.refreshEndpoint();
+      return;
+    }
     publisher?.close();
     publisher = void 0;
     publisherKey = void 0;
@@ -507,4 +513,4 @@ function startCliHqConnection(options) {
 export {
   startCliHqConnection
 };
-//# sourceMappingURL=chunk-XYKYXFEE.js.map
+//# sourceMappingURL=chunk-TE3Q32JP.js.map