@staix/agent-hub 0.12.20 → 0.12.22

package/src/cli/main.ts

@@ -1,4 +1,5 @@
 #!/usr/bin/env bun
+import { isPermissionMode } from "../hub/permission-mode.ts";
 import { currentRouting } from "../hub/routing.ts";
 import { spawn, spawnSync } from "node:child_process";
 import { chmodSync, existsSync, readFileSync, renameSync, writeFileSync } from "node:fs";
@@ -10,22 +11,26 @@ import { loadConfig } from "../hub/daemon.ts";
 import { factsHook } from "./facts-hook.ts";
 import { projectContext, realPath } from "../hub/project.ts";
 import { Registry, type Project } from "../hub/registry.ts";
-import { inspectProject, processAlive, startProject, stopProject, runProjectDaemon } from "../hub/lifecycle.ts";
+import { inspectProject, startProject, stopProject, runProjectDaemon } from "../hub/lifecycle.ts";
+import { processLiveness, processSignature } from "../pi/process-signature.ts";
 import { openManager, startManager, stopManager } from "../hub/manager.ts";
 import { OmniRoute } from "../omniroute/client.ts";
 import { MemoryClient } from "../memory/client.ts";
 import { init, planInit } from "./init.ts";
 import { launcherPreview } from "./preview.ts";
-import { buildLaunch, claudeObservationHooks, nativeLaunchEnv, UNATTENDED_WARNING } from "./launch.ts";
+import { cleanupStaleClaudeSettings, cleanupClaudeSettings, buildLaunch, claudeObservationHooks, nativeLaunchEnv, recordClaudeLaunch, runClaudeLaunch, type ClaudeLaunchRecord, UNATTENDED_WARNING } from "./launch.ts";
 import { nextStep, parseList, pluginState, type InstalledPlugin, type Marketplace } from "./setup.ts";
 import { CLASSES } from "../hub/board.ts";
 import { VERSION } from "../version.ts";
 import { freeText } from "./free-text.ts";
 import { createInterface } from "node:readline/promises";
-import { activeOperation, assertLifecycleAvailable, readOperation, recoveryLock, recoveryRunner } from "../hub/recovery-store.ts";
+import { createInterface as lineReader } from "node:readline";
+import { activeOperation, assertLifecycleAvailable, operationIdPattern, readOperation, recoveryLock, recoveryRunner, signedRunner, stopSignedRunner } from "../hub/recovery-store.ts";
 import { childEnv } from "../hub/child-process.ts";
 import { abortRecovery, createOperation, disposeRecovery, liveProjects, nextActionsText, publicOperation, recoveryCommand, registeredProjects, runRecovery, type RecoveryOperation } from "./upgrade.ts";
-import { makeRecoveryDriver, makeUpgradePlan, preserveSource } from "./upgrade-runtime.ts";
+import { endPlannedPeer, makeRecoveryDriver, makeUpgradePlan, preserveSource } from "./upgrade-runtime.ts";
+import { exactVersion, latestRelease, newerVersion } from "./recovery-package.ts";
+import { operationScreen, planScreen, type ScreenIO, type UpgradeHost } from "./upgrade-interactive.ts";
 import { recordTerminalLaunch } from "./terminal-recovery.ts";
 import { ensureMlx, inspectMlx, stopMlx } from "../models/mlx.ts";
 import { setupOllamaModel } from "./models-setup.ts";
@@ -33,13 +38,16 @@ import { setupOllamaModel } from "./models-setup.ts";
 import { unknownContext } from "../hub/context-window.ts";
 import { backendLine, contextLine, peerLine, type BackendRow, type PeerRow } from "./status-lines.ts";
 import { parseSince, readEvents } from "../hub/events.ts";
+import { armRuns, benchCompare, benchCsv, benchReport, formatCompare, formatReport as formatBenchReport, formatRuns, readRuns, runSummary, sameSuite } from "../hub/bench.ts";
+import { benchPreflight, defaultIo, runBench } from "./bench.ts";
+import { appendRecords, formatResearch, LABELS, labelTarget, RESEARCH_SCHEMA, projectKey, readStores, researchReport, taskRecords, toCsv, type Label, type TaskRecord } from "../hub/research.ts";
 import { formatReport, summarize, formatTaskReport, summarizeByTask } from "../hub/report.ts";
 import { hasTree, planUndo, repoOf, restore, Turns } from "../hub/snapshots.ts";
 import { pathWarnings } from "../hub/conflicts.ts";
 import { classifyPeerCommand, cliCommandLabel, detectCliIdentity, peerCommandRefusal } from "./identity.ts";
 import { recordCliAudit } from "./identity-audit.ts";
 import { runConsole } from "./console.ts";
-import { resolveColor } from "./console-state.ts";
+import { permissionBoundary, resolveColor } from "./console-state.ts";
 import { renderHelp } from "./help.ts";
 import { renderTailEvent } from "./tail-render.ts";
 import { archiveProblem, archiveState, damagedState, failureText, MANIFEST, planReset, resetLines, resetRuntime, startState, type ResetPlan } from "./reset.ts";
@@ -97,6 +105,7 @@ function audit(outcome: "run" | "refused" | "invalid"): void {
 if (identity.role === "invalid") { audit("invalid"); fail(`${identity.reason}; use ahub console or a terminal with no agent markers`); }
 if (identity.role === "tools" && commandAccess === "console") { audit("refused"); fail(peerCommandRefusal(identity.peer, commandLabel)); }
 if (identity.role === "tools") audit("run");
+const invokedFrom = process.cwd(); // relative paths a person types (a bench suite) are theirs, not the project's
 try { process.chdir(cwd); } catch { fail(`project directory is unavailable: ${cwd}`); }
 const unattendedEnv = process.env.AGENTHUB_UNATTENDED === "1";
 const lifecycle = { inspectProject, startProject, stopProject };
@@ -108,6 +117,13 @@ function registeredProject(): Project {
   finally { registry.close(); }
 }
 
+/** #247: this project's registration for its research records; writing needs `research.enabled` in its config. */
+function researchProject(writes = true): Project {
+  const project = matchingProject() ?? fail("this project has no hub registration; run ahub up here once first");
+  if (writes && !loadConfig(cwd).research.enabled) fail('research records are off for this project; set "research": { "enabled": true } in .agenthub/config.json first');
+  return project;
+}
+
 /** The registration of exactly this root and state directory, the only one kill and reset act on; never registers. */
 function matchingProject(): Project | undefined {
   const registry = new Registry();
@@ -124,10 +140,13 @@ function resetLockFree(): void {
 /** A manifest whose daemon is alive or uncertain, by inspectProject's rule; a crashed daemon's leftovers do not count. */
 function liveManifest(): boolean {
   const control = readControl(stateDir);
-  if (control && processAlive(control.pid) !== false) return true;
+  if (control && processLiveness(control.pid, control.pidSignature) !== "gone") return true;
+  // A signed manifest is this release's daemon, which writes no hub.pid: one beside it is an older run's leftover.
+  if (control?.pidSignature) return false;
   let text: string;
+  // Only daemons up to 0.12.20 wrote hub.pid (#226): an unsigned legacy record.
   try { text = readFileSync(join(stateDir, "hub.pid"), "utf8").trim(); } catch { return false; }
-  if (/^\d+$/.test(text)) return processAlive(Number(text)) !== false;
+  if (/^\d+$/.test(text)) return processLiveness(Number(text)) !== "gone";
   // Empty or garbled: beside a dead daemon's status it is that daemon's leftover too; alone it proves nothing.
   return !control;
 }
@@ -217,17 +236,16 @@ function hubDaemonCommand(pid: number, root: string): string | undefined {
 
 /** Live PID candidates for a registration whose project root is gone: the claim, the manifest, the legacy pid file. */
 function orphanPids(project: Project): number[] {
+  // #226: a recorded pid counts only while its signature (when one was recorded) still names its process.
   const candidates = new Set<number>();
-  if (typeof project.pid === "number") candidates.add(project.pid);
+  if (typeof project.pid === "number" && processLiveness(project.pid, project.pidSignature) === "live") candidates.add(project.pid);
   const control = readControl(project.stateDir);
-  if (control && Number.isSafeInteger(control.pid)) candidates.add(control.pid!);
+  if (control && Number.isSafeInteger(control.pid) && processLiveness(control.pid, control.pidSignature) === "live") candidates.add(control.pid!);
   try {
     const pid = Number(readFileSync(join(project.stateDir, "hub.pid"), "utf8").trim());
     if (Number.isSafeInteger(pid) && pid > 0) candidates.add(pid);
   } catch { /* no legacy pid file */ }
-  return [...candidates].filter((pid) => {
-    try { process.kill(pid, 0); return true; } catch { return false; }
-  });
+  return [...candidates].filter((pid) => processLiveness(pid) === "live");
 }
 
 const processGone = (pid: number): boolean => !processCommandLine(pid);
@@ -324,32 +342,124 @@ async function hold(t: "pause" | "resume"): Promise<void> {
   console.log(`${args[0]} is ${res.state}`);
 }
 
-function spawnRecovery(operation: RecoveryOperation): void {
+function spawnRecovery(operation: RecoveryOperation, quiet = false): void {
   const child = spawn(process.execPath, [join(operation.sourceRoot, "src/cli/main.js"), "recovery-run", operation.id], {
     cwd, detached: true, stdio: "ignore", env: { ...process.env, AGENTHUB_RECOVERY_OPERATION: operation.id },
   });
   child.on("error", () => console.error(`runner launch failed; use ${recoveryCommand(operation, "resume")}`));
   child.unref();
   // #215: the operation's own coordinator, which has every recovery command; the global ahub may be older mid-upgrade.
-  console.log(`Recovery operation ${operation.id} scheduled.\n${recoveryCommand(operation, "status")}`);
+  if (!quiet) console.log(`Recovery operation ${operation.id} scheduled.\n${recoveryCommand(operation, "status")}`);
+}
+
+/** #272: a terminal session for the interactive screens. One line reader for the whole session, in the terminal's own
+ *  line mode, so Ctrl+C stays a signal: from then on every question answers null and the screens leave. The terminal
+ *  is not read again after it (a parent that died on the same Ctrl+C may have handed it back to the shell), and no
+ *  runner is ever stopped by it. */
+async function onTerminal(screens: (io: ScreenIO) => Promise<void>): Promise<void> {
+  const rl = lineReader({ input: process.stdin, terminal: false });
+  const lines: string[] = [], waiters: ((line: string | null) => void)[] = [];
+  let ended = false, interrupted = false;
+  rl.on("line", (line) => { const next = waiters.shift(); if (next) next(line.trim()); else lines.push(line.trim()); });
+  rl.on("close", () => { ended = true; for (const waiter of waiters.splice(0)) waiter(null); });
+  const onInterrupt = () => { interrupted = true; for (const waiter of waiters.splice(0)) waiter(null); };
+  process.on("SIGINT", onInterrupt);
+  try {
+    await screens({
+      ask: (question) => {
+        if (interrupted) return Promise.resolve(null);
+        process.stdout.write(question);
+        const typed = lines.shift();
+        return typed !== undefined ? Promise.resolve(typed) : ended ? Promise.resolve(null) : new Promise((resolve) => waiters.push(resolve));
+      },
+      out: (line) => console.log(line),
+      run: (argv) => new Promise((resolve) => {
+        // The command owns the terminal while it runs; a reader whose input already ended has nothing to pause.
+        if (!ended) rl.pause();
+        const done = (code: number) => { if (!ended) rl.resume(); resolve(code); };
+        const child = spawn(process.execPath, argv, { cwd, stdio: "inherit" });
+        child.on("error", () => done(1));
+        child.on("close", (code) => done(code ?? 1));
+      }),
+      sleep: (ms) => Bun.sleep(ms),
+      interrupted: () => interrupted,
+      // Scripted input (a pipe) is answers typed ahead for the next question; only a terminal's line is a key press.
+      typed: () => !ended && lines.splice(0).length > 0,
+    });
+  } finally { process.off("SIGINT", onInterrupt); rl.close(); }
+}
+
+function upgradeHost(kind: "restart" | "upgrade", version: string): UpgradeHost {
+  const scope = kind === "restart" ? cwd : undefined;
+  return {
+    plan: () => makeUpgradePlan(kind, version, scope),
+    apply: async (plan, interrupted) => {
+      assertLifecycleAvailable();
+      const current = await makeUpgradePlan(kind, plan.version, scope);
+      if (current.fingerprint !== plan.fingerprint) throw new Error("the plan changed during review; review it again");
+      // The check above takes seconds: a Ctrl+C in that time is still a no.
+      if (interrupted()) throw new Error("interrupted; nothing was started");
+      const operation = createOperation(plan, preserveSource(plan));
+      spawnRecovery(operation, true);
+      return operation;
+    },
+    lock: recoveryLock,
+    read: (id) => readOperation<RecoveryOperation>(id),
+    runner: recoveryRunner,
+    live: (op) => liveProjects(op, makeRecoveryDriver().inspect),
+    endPeer: (planned, peer) => endPlannedPeer(planned, peer),
+    runnerStoppable: (id) => signedRunner(id) !== undefined,
+    stopRunner: (id) => stopSignedRunner(id),
+    entry: join(import.meta.dir, "main.js"),
+    version: VERSION,
+  };
 }
 
 async function upgrade(kind: "restart" | "upgrade"): Promise<void> {
   const { one, rest } = takeFlags(args, ["--to"], []);
-  if (rest.some((a) => !["--dry-run", "--yes"].includes(a)) || (kind === "restart" && one["--to"])) fail("usage: ahub upgrade --to <version> [--dry-run] [--yes] | ahub restart [--dry-run] [--yes]");
-  if (kind === "upgrade" && !one["--to"]) fail("upgrade requires --to <exact-version>");
+  if (rest.some((a) => !["--dry-run", "--yes"].includes(a)) || (kind === "restart" && one["--to"])) fail("usage: ahub upgrade [--to <version>] [--dry-run] [--yes] | ahub restart [--dry-run] [--yes]");
   if (kind === "upgrade" && selector) fail("upgrade changes the shared package/plugin; omit --project to review all affected running projects");
-  const plan = await makeUpgradePlan(kind, one["--to"] ?? VERSION, kind === "restart" ? cwd : undefined);
+  const dry = rest.includes("--dry-run"), yes = rest.includes("--yes");
+  // #272: on a terminal, with neither flag, the plan is reviewed, applied and followed in one session.
+  const interactive = !dry && !yes && !!process.stdin.isTTY && !!process.stdout.isTTY;
+  let to = one["--to"];
+  if (kind === "upgrade" && !to) {
+    if (yes || (!dry && !interactive)) fail("upgrade requires --to <exact-version>");
+    to = await latestRelease();
+    console.error(`ahub: the latest release is ${to}`);
+  }
+  // Before the value reaches a package spec or the registry: a range such as 0.13 would resolve to whatever matches.
+  if (to !== undefined) { try { exactVersion(to); } catch (error) { fail((error as Error).message); } }
+  // An older CLI must not coordinate a newer target (docs/operations.md): the target's own coordinator takes over.
+  if (kind === "upgrade" && newerVersion(to!, VERSION) && process.env.AGENTHUB_UPGRADE_HANDOVER !== "1") {
+    const argv = ["x", "--package", `@staix/agent-hub@${to}`, "ahub", "upgrade", "--to", to!, ...rest];
+    console.error(`ahub: ${to} is newer than this CLI (${VERSION}), so its own coordinator runs the upgrade: bun ${argv.join(" ")}`);
+    if (interactive) {
+      const rl = createInterface({ input: process.stdin, output: process.stdout });
+      try { if (!/^(|y|yes)$/i.test((await rl.question("Continue? [Y/n] ")).trim())) return; } finally { rl.close(); }
+    }
+    // The coordinator owns the terminal from here. This process only waits, and a Ctrl+C meant for that session must not
+    // end it first and hand the terminal back to the shell under it.
+    process.on("SIGINT", () => {});
+    const code = await new Promise<number>((resolve) => {
+      const child = spawn(process.execPath, argv, { cwd, stdio: "inherit", env: { ...process.env, AGENTHUB_UPGRADE_HANDOVER: "1" } });
+      child.on("error", (error) => { console.error(`ahub: cannot run bun x: ${error.message}`); resolve(1); });
+      child.on("close", (status) => resolve(status ?? 1));
+    });
+    process.exit(code);
+  }
+  if (interactive) return onTerminal((io) => planScreen(upgradeHost(kind, to ?? VERSION), io));
+  const plan = await makeUpgradePlan(kind, to ?? VERSION, kind === "restart" ? cwd : undefined);
   console.log(JSON.stringify(plan, null, 2));
   // #206: name every blocker and reconnect-only session on stderr, not only inside the JSON above.
   for (const p of plan.projects) for (const peer of p.reconnectOnly ?? []) console.error(`ahub: ${p.project.id}: ${peer} is reconnect-only (unmanaged session): its plugin reattaches to the new hub; no terminal is closed or relaunched`);
   for (const p of plan.projects) for (const peer of p.freshStart ?? []) console.error(`ahub: ${p.project.id}: ${peer} restarts as a new session: its thread has no rollout and the hub recorded no turn on it, so nothing is lost`);
   const blockers = [...plan.blockers, ...plan.projects.flatMap((p) => p.blockers.map((b) => `${p.project.id}: ${b}`))];
   for (const blocker of blockers) console.error(`ahub: blocker: ${blocker}`);
-  if (args.includes("--dry-run")) return;
+  if (dry) return;
   if (blockers.length) fail("plan has blockers; no runtime was changed");
   assertLifecycleAvailable();
-  if (!args.includes("--yes")) {
+  if (!yes) {
     if (!process.stdin.isTTY) fail("review --dry-run and use --yes in non-interactive sessions");
     const rl = createInterface({ input: process.stdin, output: process.stdout });
     try {
@@ -367,9 +477,18 @@ const commands: Record<string, () => Promise<void> | void> = {
   upgrade: () => upgrade("upgrade"),
   restart: () => upgrade("restart"),
   recovery: async () => {
-    const [action, id, ...rest] = args;
-    const usage = "usage: ahub recovery status|resume|abort <operation-id> | ahub recovery dispose <operation-id> --fresh-session <peer>|--stop-and-archive --reason <text>";
-    if (!id || !["status", "resume", "abort", "dispose"].includes(action ?? "") || (action !== "dispose" && rest.length)) fail(usage);
+    const [action, given, ...more] = args;
+    const usage = "usage: ahub recovery [status|resume|abort] [<operation-id>] | ahub recovery dispose [<operation-id>] --fresh-session <peer>|--stop-and-archive --reason <text>";
+    // #272: without an id, the operation that holds the machine's lock; bare on a terminal, its screen.
+    const named = !!given && operationIdPattern.test(given);
+    const id = named ? given : recoveryLock();
+    const rest = named || given === undefined ? more : [given, ...more];
+    if (action === undefined && process.stdin.isTTY && process.stdout.isTTY) {
+      if (!id) return console.log("no recovery operation is open");
+      return onTerminal(async (io) => { await operationScreen(upgradeHost("restart", VERSION), io, id); });
+    }
+    if (!["status", "resume", "abort", "dispose"].includes(action ?? "") || (action !== "dispose" && rest.length)) fail(usage);
+    if (!id) fail("no recovery operation holds the lock; name one by its id");
     const operation = readOperation<RecoveryOperation>(id);
     const runner = recoveryRunner(id);
     if (action === "status") {
@@ -525,10 +644,6 @@ const commands: Record<string, () => Promise<void> | void> = {
       process.env.AGENTHUB_INSTANCE_ID = control.instanceId;
       const terminal = await recordTerminalLaunch("claude", cwd, stateDir, control.instanceId);
       if (!terminal) process.env.AGENTHUB_LAUNCH_ID = randomUUID();
-      // Native hook identity exists in an ordinary terminal too; this is not an Orca recovery record.
-      const file = join(stateDir, "claude-launch.json");
-      writeFileSync(`${file}.tmp`, JSON.stringify({ instanceId: control.instanceId, launchId: process.env.AGENTHUB_LAUNCH_ID }), { mode: 0o600 });
-      chmodSync(`${file}.tmp`, 0o600); renameSync(`${file}.tmp`, file);
     }
     // `--settings` outranks project and user settings, so the tee has to wrap whichever status line would have won:
     // project local, then project, then user.
@@ -542,9 +657,26 @@ const commands: Record<string, () => Promise<void> | void> = {
     }
     // Turn-free facts and opted-in task sweeps share native PreToolUse/PostToolUse/Stop observations.
     const facts = claudeObservationHooks(projectConfig(), { script: join(import.meta.dir, "facts-hook.ts"), stateDir });
+    let previousLaunch: unknown;
+    try { previousLaunch = JSON.parse(readFileSync(join(stateDir, "claude-launch.json"), "utf8")); } catch { /* no previous managed launch */ }
     const launch = buildLaunch("claude", args, { unattended: unattendedEnv, statusLine: { script: join(import.meta.dir, "statusline-tee.ts"), stateDir, ...(original ? { original } : {}) }, ...(facts ? { facts } : {}) });
-    if (launch.warning) console.error(launch.warning);
-    exec(launch.cmd, launch.args, "claude");
+    let result: Awaited<ReturnType<typeof runClaudeLaunch>>;
+    try {
+      let launchRecord: ClaudeLaunchRecord | undefined;
+      if (control?.instanceId) {
+        // Publish only the final launch's installed hook and effective permission flags.
+        launchRecord = { instanceId: control.instanceId, launchId: process.env.AGENTHUB_LAUNCH_ID!, permissionHook: launch.permissionHook === true, hookPurpose: launch.hookPurpose, settingsFile: launch.settingsFile, launcherPid: process.pid, launcherSignature: processSignature(process.pid), unattended: launch.unattended === true };
+        recordClaudeLaunch(stateDir, launchRecord);
+        cleanupStaleClaudeSettings(stateDir, previousLaunch);
+      }
+      if (launch.warning) console.error(launch.warning);
+      const env = nativeLaunchEnv("claude", childEnv());
+      result = await runClaudeLaunch(launch, { cwd, stateDir, env: { ...env, AGENTHUB_STATE_DIR: stateDir, AGENTHUB_PROJECT_DIR: cwd }, ...(launchRecord ? { record: launchRecord } : {}) });
+    } finally {
+      cleanupClaudeSettings(stateDir, launch.settingsFile);
+    }
+    if (result.error) fail(`cannot run ${launch.cmd}: ${result.error.message}`);
+    process.exit(result.status ?? 1);
   },
 
   codex: async () => {
@@ -561,7 +693,7 @@ const commands: Record<string, () => Promise<void> | void> = {
     const before = readControl(stateDir);
     if (recovering && before?.instanceId) await recordTerminalLaunch("codex", cwd, stateDir, before.instanceId);
     const hub = await connect();
-    const res = await hub.request({ t: "start", peer: "codex", operationId: process.env.AGENTHUB_RECOVERY_OPERATION });
+    const res = await hub.request({ t: "start", peer: "codex", args: { unattended: launch0.unattended === true }, operationId: process.env.AGENTHUB_RECOVERY_OPERATION });
     hub.close();
     if (!res.ok) fail(res.error);
     const control = readControl(stateDir);
@@ -766,6 +898,17 @@ const commands: Record<string, () => Promise<void> | void> = {
     if (sub === "show") return console.log(await taskOp("task_show", { id: rest[0] }));
     if (sub === "escalate") return console.log(await taskOp("task_escalate", { id: rest[0] }));
     if (sub === "assign") return console.log(await taskOp("task_assign", { id: rest[0], peer: rest[1] }));
+    if (sub === "label") {
+      // #247: a person's later verdict on an approved task (the identity gate keeps it human-only), for research records.
+      const { one, rest: positional } = takeFlags(rest, ["--created"], []);
+      const id = Number(positional[0]), label = positional[1] as Label;
+      if (positional.length !== 2 || !/^[1-9]\d*$/.test(positional[0]!) || !Number.isSafeInteger(id) || !(LABELS as readonly string[]).includes(label)) fail(`usage: ahub task label <id> ${LABELS.join("|")} [--created <time from ahub research export>]`);
+      const project = researchProject();
+      let record: TaskRecord;
+      try { record = labelTarget(readStores(project.id), readEvents(join(stateDir, "events.jsonl")), id, one["--created"]); } catch (error) { fail((error as Error).message); }
+      appendRecords(project.id, [{ schema: RESEARCH_SCHEMA, kind: "label", project: projectKey(project.id), task: id, createdAt: record!.createdAt!, label, at: new Date().toISOString() }]);
+      return console.log(`task #${id} (approved ${record!.approvedAt}) labelled ${label} in the research records`);
+    }
     if (sub !== "propose" || rest.length < 1) fail("usage: ahub task propose [<class>] <title...> | show <id> | assign <id> <peer> | escalate <id>");
     // `--class` is the explicit form. A first word that is a class name is still taken as the class (the documented
     // short form), but said out loud: "review the auth module" would otherwise be filed as class review, silently.
@@ -814,6 +957,39 @@ const commands: Record<string, () => Promise<void> | void> = {
   pause: () => hold("pause"),
   resume: () => hold("resume"),
 
+  stop: async () => {
+    const peer = args[0];
+    if (args.length !== 1 || !peer || !/^[a-z][a-z0-9-]{0,31}$/.test(peer)) fail("usage: ahub stop <peer>");
+    const hub = await connect();
+    try {
+      const reply = await hub.request({ t: "peer_stop", peer }, 35_000);
+      if (reply.ok !== true) {
+        const error = String(reply.error ?? "peer stop request refused");
+        fail(/unknown (?:control )?(?:message|request|command)(?:\b|:)|this hub does not know "peer_stop"/i.test(error) ? `${error}; upgrade the running hub to use ahub stop` : error);
+      }
+      console.log(`${peer}: ${reply.state}`);
+    } finally { hub.close(); }
+  },
+
+  permission: async () => {
+    const positional = args.filter(arg => arg !== "--yes");
+    const [peer, mode] = positional;
+    if (positional.length > 2 || positional.some(arg => arg.startsWith("--")) || (mode !== undefined && !isPermissionMode(mode))) fail("usage: ahub permission [<peer> [ask|ask-when-needed|never-ask]] [--yes]");
+    if (mode === "never-ask" && !args.includes("--yes")) fail("never-ask requires --yes; nothing was changed");
+    if (mode === "never-ask" && peer) console.error(`never-ask: ${permissionBoundary(peer)}; approval prompts are disabled`);
+    const hub = await connect();
+    try {
+      const reply = await hub.request({ t: "permission", ...(peer ? { peer } : {}), ...(mode ? { mode, confirmed: args.includes("--yes") } : {}) }, 35_000);
+      if (reply.ok === false) {
+        const error = String(reply.error ?? "permission request refused");
+        fail(/unknown (?:control )?(?:message|request|command)(?:\b|:)|this hub does not know "permission"/i.test(error) ? `${error}; upgrade the running hub to use ahub permission` : error);
+      }
+      if (reply.peers) for (const [id, value] of Object.entries(reply.peers)) console.log(`${id}: ${value}`);
+      else console.log(`${peer}: ${reply.permissionMode}`);
+      if (typeof reply.note === "string" && reply.note) console.log(reply.note);
+    } finally { hub.close(); }
+  },
+
   permit: async () => {
     const [id, option] = args;
     if (!id || !option) fail("usage: ahub permit <id> <option|deny>");
@@ -831,6 +1007,7 @@ const commands: Record<string, () => Promise<void> | void> = {
     if (args.includes("--json")) return console.log(JSON.stringify(status, null, 2));
     console.log(`hub pid ${status.pid}, control 127.0.0.1:${status.controlPort}, ${status.cwd}`);
     if (status.deliveryError) console.log(`  delivery storage: ${status.deliveryError}; dispatch is stopped`);
+    for (const pending of status.permissionDefaults ?? []) console.log(`  permission default ${pending.peer}: never-ask from ${pending.source}, stays ask until a person confirms in ahub console`);
     for (const line of (status as { crash?: string[] }).crash ?? []) console.log(`  crash recovery: ${line}`);
     const peers = Object.entries(status.peers as Record<string, PeerRow>);
     for (const [id, p] of peers) console.log(peerLine(id, { ...p, context: p.context ?? unknownContext() }));
@@ -846,6 +1023,89 @@ const commands: Record<string, () => Promise<void> | void> = {
   export: () => {
     for (const e of readEvents(join(stateDir, "events.jsonl"), since())) console.log(JSON.stringify(e));
   },
+  bench: async () => {
+    // #251: benchmark suites run against the peers attached here, and their comparison across arms.
+    const json = args.includes("--json");
+    const [sub, ...rest] = args.filter((a) => a !== "--json");
+    if (sub === "run") {
+      const { one, rest: positional } = takeFlags(rest, ["--arm", "--repeat", "--tasks"], []);
+      const suite = positional[0];
+      const repeat = one["--repeat"] === undefined ? 1 : Number(one["--repeat"]);
+      if (!suite || positional.length !== 1 || !one["--arm"]?.trim() || !Number.isSafeInteger(repeat) || repeat < 1 || repeat > 100) {
+        fail("usage: ahub bench run <suite.json> --arm <label> [--repeat 1..100] [--tasks a,b]");
+      }
+      await benchPreflight(cwd, loadConfig(cwd).bench.enabled); // the suite's refs are checked once it is read
+      let interrupted = false;
+      const stop = () => { interrupted = true; console.log("stopping at the next check: the attempt in progress is recorded as interrupted (a task still open is named)"); };
+      process.on("SIGINT", stop);
+      try {
+        const run = await runBench({ cwd, stateDir, suitePath: resolve(invokedFrom, suite!), arm: one["--arm"]!.trim(), repeat, only: one["--tasks"]?.split(",").map((t) => t.trim()).filter(Boolean), stopped: () => interrupted }, defaultIo(stateDir, cwd));
+        return console.log(formatBenchReport(benchReport(readRuns().filter((r) => r.header.run === run))).join("\n"));
+      } finally { process.off("SIGINT", stop); }
+    }
+    const runs = readRuns();
+    if (sub === undefined || sub === "list") return console.log(json ? JSON.stringify(runs.map(runSummary), null, 2) : formatRuns(runs).join("\n"));
+    if (sub === "status") {
+      const live = runs.filter((r) => r.state === "running" || r.state === "unknown");
+      return console.log(json ? JSON.stringify(live.map(runSummary), null, 2) : live.length ? formatRuns(live).join("\n") : "no benchmark run is running");
+    }
+    if (sub === "report") {
+      const run = runs.find((r) => r.header.run === rest[0]) ?? fail(`usage: ahub bench report <run> (ahub bench list shows the runs)`);
+      const r = benchReport([run]);
+      return console.log(json ? JSON.stringify(r, null, 2) : formatBenchReport(r).join("\n"));
+    }
+    if (sub === "compare") {
+      const mixed = rest.includes("--mixed");
+      const { one, rest: picks } = takeFlags(rest.filter((a) => a !== "--mixed"), ["--suite"], []);
+      const scoped = one["--suite"] ? runs.filter((r) => r.header.suite === one["--suite"]) : runs;
+      const groups = picks.map((pick) => {
+        const byRun = scoped.filter((r) => r.header.run === pick);
+        const chosen = byRun.length ? byRun : armRuns(scoped, pick);
+        if (!chosen.length) fail(`no run, or arm with a run that is over, named ${pick}${one["--suite"] ? ` in suite ${one["--suite"]}` : ""}`);
+        return { arm: byRun.length ? `${pick} [${byRun[0]!.header.arm}]` : pick, runs: chosen };
+      });
+      if (groups.length < 2) fail("usage: ahub bench compare <run|arm> <run|arm>... [--suite <name>] [--mixed] [--json]");
+      const differs = sameSuite(groups);
+      if (differs && !mixed) fail(`${differs}; compare runs of one suite file (name runs by id, or --suite when the names differ), or pass --mixed to compare them anyway`);
+      const c = benchCompare(groups);
+      return console.log(json ? JSON.stringify(c, null, 2) : formatCompare(c).join("\n"));
+    }
+    if (sub === "export") {
+      const format = rest.includes("--format") ? rest[rest.indexOf("--format") + 1] : "jsonl";
+      if (format !== "jsonl" && format !== "csv") fail("--format takes jsonl or csv");
+      if (format === "csv") return void process.stdout.write(benchCsv(runs));
+      for (const r of runs) for (const record of [r.header, ...r.attempts, ...(r.end ? [r.end] : [])]) console.log(JSON.stringify(record));
+      return;
+    }
+    fail("usage: ahub bench run|list|status|report|compare|export (ahub help bench)");
+  },
+  research: () => {
+    // #247: measures from the opt-in research records; `export` writes them out, `backfill` builds them from events.jsonl.
+    const [sub] = args;
+    // Another project is `ahub --project <dir> research`; anything unknown is refused rather than ignored.
+    const flags = new Set(["--all", "--json", "--since", "--format"]), valued = new Set(["--since", "--format"]);
+    for (let i = sub === "export" || sub === "backfill" ? 1 : 0; i < args.length; i++) {
+      if (!flags.has(args[i]!)) fail(`ahub research: unknown argument ${args[i]} (another project: ahub --project <dir> research)`);
+      if (valued.has(args[i]!)) i++;
+    }
+    if (sub === "backfill") {
+      const project = researchProject();
+      const written = appendRecords(project.id, taskRecords(readEvents(join(stateDir, "events.jsonl")), project.id, { version: VERSION, source: "backfill" }));
+      return console.log(`${written} research record${written === 1 ? "" : "s"} written from this project's events`);
+    }
+    const all = args.includes("--all");
+    const records = readStores(all ? undefined : researchProject(false).id);
+    const from = since();
+    if (sub === "export") {
+      const format = args[args.indexOf("--format") + 1] ?? "jsonl";
+      if (args.includes("--format") && !["jsonl", "csv"].includes(format)) fail("--format takes jsonl or csv");
+      const kept = records.filter((r) => r.kind === "label" || Date.parse(r.approvedAt) >= from);
+      return void process.stdout.write(args.includes("--format") && format === "csv" ? toCsv(kept) : kept.map((r) => JSON.stringify(r)).join("\n") + (kept.length ? "\n" : ""));
+    }
+    if (sub !== undefined && !sub.startsWith("--")) fail("usage: ahub research [--since 30d] [--all] [--json] | export [--format jsonl|csv] [--since] [--all] | backfill");
+    const r = researchReport(records, from);
+    console.log(args.includes("--json") ? JSON.stringify(r, null, 2) : formatResearch(r).join("\n"));
+  },
   report: () => {
     const by = args.indexOf("--by");
     if (by >= 0 && args[by + 1] !== "task") return fail("--by takes task");
@@ -1020,7 +1280,7 @@ const commands: Record<string, () => Promise<void> | void> = {
     // The prefix tells a concurrent reset's claim from a daemon's.
     const claim = `reset-${randomUUID()}`;
     try {
-      if (!registry.claim(project.id, claim, process.pid)) {
+      if (!registry.claim(project.id, claim, process.pid, processSignature(process.pid))) {
         const holder = registry.get(project.id);
         throw new Error(holder?.instanceId?.startsWith("reset-") ? `another ahub reset of this project is running (pid ${holder.pid}); nothing was reset` : "a hub started after the stop; nothing was reset, run ahub reset again");
       }

package/src/cli/preview.ts

@@ -38,9 +38,19 @@ export function launcherPreview(tool: "claude" | "codex" | "kimi" | "pi", raw: s
     }
     // Build with original inputs so channel selection and owned-flag validation match normal launch.
     const facts = claudeObservationHooks(config, { script: join(import.meta.dir, "facts-hook.ts"), stateDir });
-    launch = buildLaunch(tool, args, { unattended, statusLine: { script: join(import.meta.dir, "statusline-tee.ts"), stateDir, ...(original ? { original: { command: original.command ? REDACTED : "", ...(typeof original.refreshInterval === "number" && Number.isFinite(original.refreshInterval) ? { refreshInterval: original.refreshInterval } : {}), ...(typeof original.padding === "number" && Number.isFinite(original.padding) ? { padding: original.padding } : {}) } } : {}) }, ...(facts ? { facts } : {}) });
-    const passthrough = args.filter(arg => !["--unattended", "--safe", "--new"].includes(arg));
-    launch.args = launch.args.slice(0, launch.args.length - passthrough.length).concat(passthrough.map(publicArg));
+    launch = buildLaunch(tool, args, { preview: true, unattended, statusLine: { script: join(import.meta.dir, "statusline-tee.ts"), stateDir, ...(original ? { original: { command: original.command ? REDACTED : "", ...(typeof original.refreshInterval === "number" && Number.isFinite(original.refreshInterval) ? { refreshInterval: original.refreshInterval } : {}), ...(typeof original.padding === "number" && Number.isFinite(original.padding) ? { padding: original.padding } : {}) } } : {}) }, ...(facts ? { facts } : {}) });
+    // Merged caller settings change the native argument count. Rebuild a redacted preview,
+    // retaining the managed hooks but never serializing caller settings or their file path.
+    const safeArgs = args.map((arg, i) => {
+      if (args[i - 1] === "--settings") return "{}";
+      if (arg.startsWith("--settings=")) return "--settings={}";
+      if (["--unattended", "--safe", "--new"].includes(arg)) return arg;
+      return publicArg(arg);
+    });
+    const publicLaunch = buildLaunch(tool, safeArgs, { preview: true, unattended, statusLine: { script: join(import.meta.dir, "statusline-tee.ts"), stateDir,
+      ...(original ? { original: { command: original.command ? REDACTED : "", ...(typeof original.refreshInterval === "number" && Number.isFinite(original.refreshInterval) ? { refreshInterval: original.refreshInterval } : {}), ...(typeof original.padding === "number" && Number.isFinite(original.padding) ? { padding: original.padding } : {}) } } : {}) }, facts });
+    publicLaunch.args.splice(0, 2, ...launch.args.slice(0, 2)); // channel selection was verified against the original MCP config
+    launch = publicLaunch;
   } else if (tool === "codex") {
     // A fresh native proxy is assigned by start, never guessed from an old status file.
     launch = buildLaunch(tool, args, { unattended, proxyUrl: UNRESOLVED, codexBin: config.codex_bin === "codex" ? "codex" : REDACTED });

package/src/cli/recovery-package.ts

@@ -67,6 +67,23 @@ export async function registryRelease(version: string, run: RunCommand = runComm
   return { version, integrity: data["dist.integrity"] };
 }
 
+/** #272: the registry's latest release, the target when `--to` is left out. */
+export async function latestRelease(run: RunCommand = runCommand): Promise<string> {
+  const result = await run(["npm", "view", "@staix/agent-hub", "version", "--json"], { timeoutMs: 30_000 });
+  if (result.code !== 0) throw new Error("registry metadata unavailable for @staix/agent-hub; name the release with --to <version>");
+  let version: unknown;
+  try { version = JSON.parse(result.stdout); } catch { /* reported below */ }
+  if (typeof version !== "string") throw new Error("registry returned an unexpected latest release; name the release with --to <version>");
+  return exactVersion(version);
+}
+
+/** ponytail: compares the numeric triple only, so a pre-release never counts as newer than its own release; a semver compare if pre-releases are ever published. */
+export function newerVersion(a: string, b: string): boolean {
+  const [x, y] = [a, b].map((v) => v.split("-")[0]!.split(".").map(Number));
+  for (let i = 0; i < 3; i++) if (x![i] !== y![i]) return (x![i] ?? 0) > (y![i] ?? 0);
+  return false;
+}
+
 /** Stage outside the mutable global install; retain successful versions for recovery. */
 export async function stageRelease(version: string, integrity: string, run: RunCommand = runCommand, home = hubHome()): Promise<{ root: string; digest: string }> {
   exactVersion(version);

package/src/cli/status-lines.ts

@@ -1,7 +1,7 @@
 import type { ContextView } from "../hub/context-window.ts";
 // What `ahub status` prints for a peer and for a model backend, kept pure so it can be checked.
 
-export interface PeerRow { context?: ContextView; state?: string; queued?: number; queuedImportant?: number; needsReview?: number; heldBy?: string; holdNote?: string; liveAccepted?: string[]; oldestQueuedAt?: number; attached?: boolean; toolsOnly?: string; paused?: string; servedBy?: string; requestedModel?: string }
+export interface PeerRow { permissionMode?: string; context?: ContextView; state?: string; queued?: number; queuedImportant?: number; needsReview?: number; heldBy?: string; holdNote?: string; liveAccepted?: string[]; oldestQueuedAt?: number; attached?: boolean; toolsOnly?: string; paused?: string; servedBy?: string; requestedModel?: string }
 export interface BackendRow { kind?: string; alias?: string; state?: string; active?: number; requestedModel?: string; actualModel?: string; provider?: string; coolingUntil?: string; failures?: number; failingUntil?: string }
 
 /** Aliases are already namespaced ("dgx/coding"); only an alias that is not gets its kind in front of it. */
@@ -26,6 +26,7 @@ export function peerLine(id: string, p: PeerRow): string {
     (p.oldestQueuedAt !== undefined ? `  oldest ${Math.max(0, Math.floor((Date.now() - p.oldestQueuedAt) / 1000))}s` : "") +
     (p.attached === false ? "  disconnected" : "") +
     (p.toolsOnly ? `  ${p.toolsOnly}` : "") +
+    (p.permissionMode && p.permissionMode !== "ask" ? `  permission: ${p.permissionMode}` : "") +
     (p.paused ? `  (${p.paused})` : "") +
     (p.servedBy ? `  last call: ${p.servedBy}` : "") +
     // Which backend the peer asked for on its last turn: without it, telling a Pi DGX turn from an MLX one