@staix/agent-hub 0.24.0 → 0.26.0

package/src/hub/daemon.ts

@@ -150,6 +150,7 @@ import { closeSync, constants as fsConstants, openSync, lstatSync, readSync, sta
 import { basename, isAbsolute } from "node:path";
 import { fileURLToPath } from "node:url";
 import { canonicalPath, realPath, hubHome } from "./project.ts";
+import { HandoffRefusal, checkTarget, parseHandoff, resolveTarget, handoffEventFor, type HandoffEvent } from "./handoff.ts";
 import { mainCheckoutBranch, mainCheckoutWarning } from "./main-checkout.ts";
 import type { BusEvent, PublishResult } from "./bus.ts";
 import { CONDUCTOR_TOOLS, CONDUCTOR_TOOL_NAMES, DEFAULT_ROLES, IMPLEMENTER_HUB_TOOL_NAMES, roleContract, TASK_TOOL_NAMES, TASK_TOOLS, BOUNDED_TASK_TOOLS, CAPABILITY_TOOLS, profileTaskTools } from "./hub-tools.ts";
@@ -173,15 +174,20 @@ import { githubOriginRepository, githubRemoteRepository } from "./task-repositor
 import { Bus } from "./bus.ts";
 import { attribute, deliveryTask, type TaskAttribution } from "./attribution.ts";
 import { DeliveryJournal } from "./delivery-journal.ts";
+import { assertHubDbCompatible } from "./db-schema.ts";
 import { startDashboard, dashboardView, type DashboardSession } from "./ui.ts";
 import { codeGroup } from "./task-code-group.ts";
-import { DAEMON_FEATURES, PROTOCOL, readControl, stateDirFor } from "./control-client.ts";
+import { ControlClient, DAEMON_FEATURES, PROTOCOL, PROTOCOL_MIN, readControl, stateDirFor } from "./control-client.ts";
+import { negotiateProtocol, protocolWindow } from "./protocol-window.ts";
+import { attachEvent, originOf, type Origin } from "./provenance.ts";
+import { mintSessionId, sessionGate } from "./session-identity.ts";
 import { MAX_HOP, newEnvelope, parseMarker, replyParent, sanitize, USER, type Envelope, type PeerId, type Priority } from "./envelope.ts";
 import { BasePeer, DEFAULT_WATCHDOG_MS, type FrontendProxyPort, type PeerAdapter } from "./peers.ts";
 import { ClaudeTodoSource, DoingTracker, readAiTitle, type DoingKind, type PeerDoing, type TitleCursor } from "./peer-doing.ts";
 import { MemoryClient } from "../memory/client.ts";
 import { VERSION } from "../version.ts";
 import { appendRecords, taskRecords } from "./research.ts";
+import { parseResearchRunners, type RegisteredRunnerV1 } from "./research-runner.ts";
 import { benchReport, MIN_ATTEMPTS, readRuns, runSummary, type Measures, type RunSummary } from "./bench.ts";
 import { projectChain } from "@staix/agent-mem";
 import { recallFor } from "../memory/recall.ts";
@@ -196,6 +202,7 @@ import { consumeRecoveryContinuity, readRecoveryContinuity, retireRecoveryContin
 import { shellQuote, publishTerminalLaunch, terminalLaunchCandidate, launcherOf } from "../cli/terminal-recovery.ts";
 import { TerminalLaunches, terminalOwnerReads } from "./terminal-launch.ts";
 
+import { START_AGENTS_DEFAULT, checkStartAgents } from "./start-agents.ts";
 import { peerStartConfig, PEER_START_RETRY_MS, startModeOf, terminalTemplate, type PeerStartConfig, type StartMode } from "./start-mode.ts";
 import { hasKimiHubMcp } from "../cli/kimi-mcp.ts";
 import type { TerminalOpener } from "./terminal-open.ts";
@@ -278,13 +285,15 @@ export interface HubConfig {
   /** A git tree at each turn boundary for `ahub turns` and `ahub undo`, the last `keep` per peer (issue #33). */
   snapshots: { enabled: boolean; keep: number };
   /** #247: opt-in research records of each approved task in ~/.agenthub/research (ids, counts and tokens only). */
-  research: { enabled: boolean };
+  research: { enabled: boolean; runners?: Record<string, RegisteredRunnerV1> };
   /** #251: a project kept for benchmarks: `ahub bench run` resets its work tree between attempts. */
   bench: { enabled: boolean };
   /** #269: per peer, how a start comes up when nothing more specific was asked: `tui` (the default) or `headless`. Pi and Codex only. */
   peers: PeerStartConfig;
   /** #269: a machine-local argv that opens a terminal for a hub-made TUI start; `{command}` in one element is the command to run. */
   terminal: { open: string[] };
+  /** #608: the peers `ahub up` asks the hub to start, machine-local. Read by the CLI at a person's `ahub up` and by the dashboard manager's start; the daemon itself never starts one from it. */
+  start: { agents: string[] };
   /** Per-sender rate limits and repeat suppression for what agents send (issue #38). */
   limits: LimitsConfig;
   /** Reviewer choice from recorded review outcomes, once a reviewer has `min_reviews` of an implementer (issue #35). */
@@ -350,6 +359,7 @@ export const DEFAULT_CONFIG: HubConfig = {
   bench: { enabled: false },
   peers: {},
   terminal: { open: [] },
+  start: { agents: [...START_AGENTS_DEFAULT] },
   limits: DEFAULT_LIMITS,
   review: { adaptive: false, min_reviews: 5 },
   recovery: { auto_resume_after_crash: false, reconnect_wait_s: RECOVERY_RECONNECT_WAIT_SETTING.defaultValue, unmanaged_reconnect_wait_s: RECOVERY_UNMANAGED_RECONNECT_WAIT_SETTING.defaultValue, control_timeout_s: RECOVERY_CONTROL_TIMEOUT_SETTING.defaultValue, shutdown_wait_s: RECOVERY_SHUTDOWN_WAIT_SETTING.defaultValue, terminal_idle_wait_s: RECOVERY_TERMINAL_IDLE_WAIT_SETTING.defaultValue },
@@ -360,14 +370,29 @@ export const DEFAULT_CONFIG: HubConfig = {
 export { stateDirFor };
 
 /** Ids an external process may not claim: the console user and the adapters the daemon runs itself. */
-const READ_ONLY_HUB_TOOLS = new Set(["hub_capabilities", "hub_inbox", "hub_send", "hub_delivery_done", "hub_review", "hub_task_show", "hub_task_list", "hub_memory_search", "hub_memory_show", "hub_memory_graph", "hub_source_read", "hub_status"]);
+const READ_ONLY_HUB_TOOLS = new Set(["hub_capabilities", "hub_inbox", "hub_send", "hub_delivery_done", "hub_review", "hub_task_show", "hub_task_list", "hub_study_show", "hub_memory_search", "hub_memory_show", "hub_memory_graph", "hub_source_read", "hub_status"]);
 const BASE_RESERVED_IDS = new Set<string>([USER, ...SUPPORTED_PEERS.filter(peer => definitionOf(peer)?.attach?.externalHello !== true), HUB, DIGEST]);
 const PEER_ID = /^[a-z][a-z0-9-]{0,31}$/;
 
 /** The shared project config, then the machine's own file, which overrides it block by block (issue #17). */
 const CONFIG_FILES = INSTANCE_CONFIG_FILES;
 export interface ConfigLayer { name: typeof CONFIG_FILES[number]; raw: Record<string, any> }
-const CONFIG_BLOCKS = ["action_policy", "memory", "roster", "roles", "catalog_roles", "pii", "data", "conductor", "budget", "context", "inference", "omniroute", ...NATIVE_CONFIG_BLOCKS, "approvals", "tasks", "task_sweep", "bus", "adapter_timing", "model_relay", "console", "corpus", "checks", "snapshots", "research", "bench", "limits", "review", "recovery", "capabilities", "mlx", "permission_modes", "peers", "terminal", "pet"];
+const CONFIG_BLOCKS = ["action_policy", "memory", "roster", "roles", "catalog_roles", "pii", "data", "conductor", "budget", "context", "inference", "omniroute", ...NATIVE_CONFIG_BLOCKS, "approvals", "tasks", "task_sweep", "bus", "adapter_timing", "model_relay", "console", "corpus", "checks", "snapshots", "research", "bench", "limits", "review", "recovery", "capabilities", "mlx", "permission_modes", "peers", "terminal", "start", "pet"];
+
+/** Older targets parse raw keys before placement/trust checks, so even an empty or refused declaration needs the floor. */
+function rawResearchRunnersPresent(cwd: string): boolean {
+  const texts = configSourceTexts(cwd);
+  let present = false;
+  for (const name of CONFIG_FILES) {
+    if (texts[name] === null) continue;
+    const raw: unknown = JSON.parse(texts[name]!);
+    if (!raw || typeof raw !== "object" || Array.isArray(raw)) throw new Error("invalid configuration");
+    const research = (raw as Record<string, unknown>).research;
+    if (research !== undefined && (!research || typeof research !== "object" || Array.isArray(research))) throw new Error("invalid research configuration");
+    present ||= research !== undefined && Object.hasOwn(research, "runners");
+  }
+  return present;
+}
 
 /** Connection-time policy refresh reads role/feed fields only, never launches or machine-local configuration. */
 export function loadConductorPolicy(cwd: string, layers?: readonly ConfigLayer[], scope: Pick<ReturnType<typeof readInstanceConfig>, "instances" | "catalog"> = readInstanceConfig(cwd, layers)): { roles: Record<string, string[]>; conductor: HubConfig["conductor"]; planner?: string | string[]; notices?: string[] } {
@@ -475,6 +500,8 @@ export function loadConfig(cwd: string, layers?: readonly ConfigLayer[], routing
     peerStartConfig(file.peers, instances, adapterCatalog);
     if (file.terminal != null && (typeof file.terminal !== "object" || Array.isArray(file.terminal))) throw new Error("terminal must be an object");
     terminalTemplate(file.terminal?.open);
+    if (file.start != null && (typeof file.start !== "object" || Array.isArray(file.start) || Object.keys(file.start).some((key: string) => key !== "agents"))) throw new Error("start must be an object with an agents list");
+    if (file.start?.agents !== undefined) checkStartAgents(file.start.agents);
     if (file.pet !== undefined) validatePetConfig(file.pet, "config.pet");
     if (file.pet && typeof file.pet === "object" && !Object.keys(file.pet).length) delete file.pet;
     delete file.declaredExternalInstances; // host-derived refusal rows, never a file claim
@@ -484,8 +511,11 @@ export function loadConfig(cwd: string, layers?: readonly ConfigLayer[], routing
       if (name !== "config.local.json" || configTracked(cwd, name) !== false) throw new Error("remote checks require untracked config.local.json");
       if (file.checks.remote !== undefined) remoteRunnerConfig(file.checks.remote);
     }
+    if (file.research !== undefined && (!file.research || typeof file.research !== "object" || Array.isArray(file.research))) throw new Error("research must be an object");
+    const hasRunners = file.research && Object.hasOwn(file.research, "runners"), runnerDeclarations: unknown = hasRunners ? file.research.runners : undefined;
     const why = stripUntrusted(file, DEFAULT_CONFIG, cwd, name);
     if (why) ignored.push(why);
+    if (hasRunners) file.research.runners = parseResearchRunners(runnerDeclarations);
     for (const [peer, mode] of Object.entries(permissionDefaults(file.permission_modes, instances, adapterCatalog))) appliedPermissionSources[peer] = { mode, source: `.agenthub/${name}` };
     return [file];
   });
@@ -572,10 +602,11 @@ export function loadConfig(cwd: string, layers?: readonly ConfigLayer[], routing
     checks: { ...DEFAULT_CONFIG.checks, ...file.checks, runner: file.checks?.runner === "remote" ? "remote" : file.checks?.runner === "container" ? "container" : undefined,
       ...(file.checks?.remote !== undefined || file.checks?.runner === "remote" ? { remote: remoteRunnerConfig(file.checks?.remote) } : {}) },
     snapshots: { ...DEFAULT_CONFIG.snapshots, enabled: true, ...file.snapshots },
-    research: { enabled: file.research?.enabled === true }, // opt-in: anything but true is off
+    research: { enabled: file.research?.enabled === true, ...(file.research?.runners !== undefined ? { runners: parseResearchRunners(file.research.runners) } : {}) }, // opt-in: anything but true is off
     bench: { enabled: file.bench?.enabled === true }, // opt-in: a run resets this project's work tree
     peers: { ...instanceModes, ...peerStartConfig(file.peers, instances, adapterCatalog) },
     terminal: { open: terminalTemplate(file.terminal?.open) },
+    start: { agents: file.start?.agents === undefined ? [...START_AGENTS_DEFAULT] : checkStartAgents(file.start.agents) },
     limits: { ...PROJECT_LIMITS, ...file.limits }, // on with any project config (issue #38)
     review: { ...DEFAULT_CONFIG.review, ...file.review },
     recovery: { ...DEFAULT_CONFIG.recovery, ...file.recovery },
@@ -695,8 +726,11 @@ export interface DaemonOptions {
 
 interface Client {
   authed: boolean;
+  protocol?: number;
   role?: "peer" | "console" | "tools";
   peer?: PeerId;
+  /** agent-hub#1036: this connection is a distinct session; `peer` is the id the daemon minted for it (`<peer>~<origin>`), never a bus peer. */
+  session?: true;
   launchId?: string;
   externalOwner?: { owner: PeerAdapter; launchId: string; generation?: string };
   toolsOwner?: { owner: PeerAdapter; launchId: string; generation: string; cwd: string };
@@ -742,6 +776,8 @@ function capturedRecoverySourceDigest(): string | undefined {
 }
 
 export async function startDaemon(opts: DaemonOptions) {
+  // Refuse unknown formats before logs, session cleanup, module factories or any writable store is opened.
+  assertHubDbCompatible(join(opts.stateDir, "hub.db"));
   const resolveExecutionCwd = (peer: string, currentDeclarations: InstancesConfig | undefined, currentRoles: readonly string[], catalog: ReturnType<typeof createProjectAdapterCatalog>): string =>
     resolveInstanceExecutionCwd(opts.cwd, peer, currentDeclarations, currentRoles, catalog, undefined, undefined, opts.instanceExecutionGit);
   const projectId = opts.projectId ?? randomUUID();
@@ -805,6 +841,27 @@ export async function startDaemon(opts: DaemonOptions) {
   const childToolAuthority = new AsyncLocalStorage<{ by: string; current: () => boolean }>();
   const implementerTranscripts = new Map<string, string>();
   const executionOwners = new Map<string, { owner: PeerAdapter; cwd: string; socket?: Sock; launchId?: string; generation?: string | number }>();
+  /** #604 PR B: envelope id to the connection that sent it, written where a connection builds an envelope and read once when the envelope event is logged. Bounded: an entry whose event never comes ages out. */
+  const envelopeOrigins = new Map<string, Origin>();
+  /** agent-hub#1036: the ids this daemon minted for live distinct sessions. Only the daemon makes them, so `PEER_ID` is not widened and no `@` reaches the remote-room paths. */
+  const sessionIds = new Set<string>();
+  const rememberOrigin = (id: string, origin: Origin): void => { envelopeOrigins.set(id, origin); if (envelopeOrigins.size > 1024) envelopeOrigins.delete(envelopeOrigins.keys().next().value as string); };
+  const takeOrigin = (id: string): Origin | undefined => { const origin = envelopeOrigins.get(id); envelopeOrigins.delete(id); return origin; };
+  /**
+   * agent-hub#1036: the origin of an envelope that no control connection sent, one an adapter published itself (a native peer's final answer goes through `peer.onMessage`, not the `send` case).
+   * It names the sending peer and, where the adapter knows one, the launch it runs under: a channel session's `managedLaunchId`, an owned Jcode's launch id. The hub's own writers (`hub`, `user`,
+   * `digest`) and remote room ids have none, as before. Evidence for the log, never a path or text.
+   */
+  const adapterOrigin = (env: Envelope): Origin | undefined => {
+    if (env.from === "hub" || env.from === USER || env.from === "digest" || env.from.includes("@")) return undefined;
+    const peer = bus.peers.get(env.from);
+    if (!peer) return undefined;
+    // Read by shape, not by a native-branch predicate: a channel session reports `managedLaunchId`, an owned Jcode `ownedLaunchId`, and the value only labels the log. Nothing here decides
+    // authority or ownership, so it is not an `isWsPeer` host branch (the #625 ratchet lists those, and this is not one).
+    const evidence = peer as { managedLaunchId?: unknown; ownedLaunchId?: unknown };
+    const launchId = [evidence.managedLaunchId, evidence.ownedLaunchId].find((value): value is string => typeof value === "string" && value.length > 0);
+    return originOf({ role: "peer", peer: env.from, ...(launchId ? { launchId } : {}) });
+  };
   const externalLaunches = new Map<string, { owner: PeerAdapter; options: AcpOptions; launchId: string; binding: ExternalInstanceBinding; workspace: string; generation?: string }>();
   const externalBindings = new Map<string, ExternalInstanceBinding>();
   const externalSessions = new Map<string, string>();
@@ -1154,6 +1211,7 @@ export async function startDaemon(opts: DaemonOptions) {
         for (const request of sock.data.invocationRequests?.values() ?? []) request.abort();
         sock.data.invocationRequests?.clear();
         nativeHookLaunches.endOwner(sock);
+        if (sock.data.session && sock.data.peer) sessionIds.delete(sock.data.peer);
         consoles.delete(sock);
         operationalLogSubscribers.delete(sock);
         sock.data.tail?.();
@@ -1353,6 +1411,18 @@ export async function startDaemon(opts: DaemonOptions) {
     onImportantQueued: (peer, counts) => { const p = bus.peers.get(peer); if (isWsPeer(p)) p.notifyQueued(counts); } });
   startupCleanup.push(() => bus.closeJournal());
   const claudeLaunchProof = createClaudeLaunchProof({ stateDir: opts.stateDir, cwd: opts.cwd, instanceId, definitionOf, peer: peer => bus.peers.get(peer) });
+  /**
+   * agent-hub#1036: whether a tools hello is a distinct session instead of the peer it names. True when the plugin asked for it (`session`: it has no project of its own and inherited this hub's
+   * state), or when the name is a Claude channel peer and the hello PRESENTS a launch id that is not the current launch record's. An ABSENT id is never enough: the facts hook, a native
+   * helper or a person's own shell connects as the peer with none, and it must keep working (agent-hub#1036 AC4). Everything else keeps its admission as before: codex, kimi, qwen, agent
+   * shells without an inherited id, owned ACP peers and the Claude implementer child are not affected.
+   */
+  const sessionHello = (msg: Record<string, any>): boolean => {
+    if (typeof msg.session === "string") return true;
+    const peer = String(msg.peer ?? "");
+    if (definitionOf(peer)?.transport !== "claude-channel" || definitionOf(peer)?.start.daemonPort === "claude-implementer") return false;
+    try { const launch = readClaudeLaunchProof(opts.stateDir, peer, instanceId); return !!launch && typeof msg.launchId === "string" && msg.launchId !== launch.launchId; } catch { return false; }
+  };
   function frontendProxyState(peer: string, owner: PeerAdapter | undefined): { present: boolean; port: FrontendProxyPort | undefined } {
     if (!owner || !isFixedBuiltinDefinition(definitionOf(peer)) || !("frontendProxy" in owner)) return { present: false, port: undefined };
     return { present: true, port: owner.frontendProxy };
@@ -2009,6 +2079,12 @@ export async function startDaemon(opts: DaemonOptions) {
   };
   const tasks: Tasks = new Tasks({
     releaseWindow,
+    researchRunners: () => {
+      if (config.research.runners === undefined) return {};
+      const why = configRefusal(opts.cwd, "config.local.json");
+      if (why || configSourceTexts(opts.cwd)["config.local.json"] === null) throw new Error(`research.runners source trust unavailable${why ? `: ${why}` : ""}`);
+      return parseResearchRunners(config.research.runners);
+    },
     capabilities: peerCapabilities, repository: memoryRepo,
     contextAdmission: () => ({
       policy: { max_used: config.context.max_assignment_used ?? 0, stale_min: config.context.stale_min },
@@ -2140,6 +2216,8 @@ export async function startDaemon(opts: DaemonOptions) {
     // #334: the hub model's split, validated against the same `hub_backlog_split` parts schema a planner's call is, and the source registry for `--source`.
     intake: { split: (system, user, opts) => inference?.split(system, user, opts) ?? Promise.resolve(undefined), sourceKnown: id => sourceRegistry.get(id) !== undefined,
       partsSchema: () => profileTaskTools(resolvedProfile, lifecycleConfig.stages, lifecycleConfig.mode !== "off").find(tool => tool.name === "hub_backlog_split")?.inputSchema.properties.parts },
+    studySource: id => sourceRegistry.get(id),
+    experimentCheck: run => checkRunner.result(run),
     // The PII screen (issue #198): the model on this machine when MLX/Ollama is enabled, else the gateway on campus only.
     piiScreen: opts.piiScreen ?? ((text) => screenPii(text, { ...(config.mlx.enabled === false ? {} : { device: () => ensureMlx(config.mlx) }), onCampus, omni, fixedModel: () => activeRouting().local.fixed_model })),
     recordScreen: (record) => event({ type: "pii_screen", ...record }),
@@ -3166,6 +3244,71 @@ export async function startDaemon(opts: DaemonOptions) {
   // commit waits for those in flight, or its integrity digest misses their later writes. Completion checks outlive
   // their op, so recoveryReady() also waits for `tasks.checksPending()`: a check the commit's stop kills would write.
   let taskOpsInFlight = 0;
+  /**
+   * #604 ORIGIN side. Order matters: the caller's authority and the PII gates run first (nothing has left this project), then the target is resolved from the machine registry and its own
+   * status descriptor is read, and only then is the request sent. The target answers with the task id it created, which is recorded here. A refusal at any step leaves no row in either board.
+   */
+  async function handoffOut(by: PeerId, input: unknown): Promise<{ project: string; task: number; state: string; note?: string }> {
+    const request = parseHandoff(input);
+    const own = request.task === undefined ? undefined : board.get(request.task);
+    if (request.task !== undefined && !own) throw new HandoffRefusal("invalid", `no task #${request.task} in this project to hand off from`);
+    tasks.handoffOutAuthority(by, own);
+    const self = { id: projectId, root: opts.cwd };
+    const target = resolveTarget(request.project, registeredProjects(hubHome()).map(project => ({ id: project.id, root: project.root, stateDir: project.stateDir })), self, canonicalPath);
+    const descriptor = readControl(target.stateDir);
+    checkTarget(target, descriptor && { ...(descriptor.features ? { features: descriptor.features } : {}), ...(descriptor.projectId ? { projectId: descriptor.projectId } : {}) });
+    let client: ControlClient;
+    try { client = await ControlClient.connect(target.stateDir, { role: "console", projectRoot: descriptor!.cwd ?? target.root }, 5000); }
+    catch { throw new HandoffRefusal("not_running", `the hub of ${target.id} did not answer; run \`ahub up\` in ${target.root}, then hand the task off again (handoffs are not queued)`); }
+    try {
+      const { project: _project, task: originTask, ...body } = request;
+      const reply = await client.request({ t: "task_handoff", origin: { projectId, peer: by, ...(originTask === undefined ? {} : { taskId: originTask }) }, request: body }, 30_000);
+      if (!reply.ok) throw new HandoffRefusal("invalid", sanitize(String(reply.error ?? "the target refused the handoff")).slice(0, 300));
+      const remote = Number(reply.task);
+      if (!Number.isSafeInteger(remote) || remote < 1) throw new HandoffRefusal("invalid", "the target answered without a task id");
+      tasks.handoffSent(by, { projectId: target.id }, remote, request.title, request.task);
+      return { project: target.id, task: remote, state: String(reply.state ?? "proposed"), ...(typeof reply.note === "string" && reply.note ? { note: sanitize(reply.note).slice(0, 300) } : {}) };
+    } finally { client.close(); }
+  }
+  /** The head a handoff reports: the owner's `refs.commit`, else the receipt's verified head. Only a full 40-hex SHA crosses; anything else is dropped. */
+  const handoffHead = (task: Task): string | undefined => { const head = task.refs.commit ?? task.sdlc?.head ?? task.sdlc?.localHead; return typeof head === "string" && /^[0-9a-f]{40}$/.test(head) ? head : undefined; };
+  /** #604 TARGET side: tell the origin project's daemon about a state change of a task it handed us. Best effort and never blocking a board write: an origin that is down reads the state later. */
+  const handoffNotifyOrigin = (task: Task, entry: { event: string }): void => {
+    const event: HandoffEvent | undefined = handoffEventFor(entry.event, task.state);
+    const origin = event ? board.handoffOrigin(task.id) : undefined;
+    if (!event || !origin) return;
+    void (async () => {
+      try {
+        const project = registeredProjects(hubHome()).find(candidate => candidate.id === origin.projectId);
+        const descriptor = project && readControl(project.stateDir);
+        if (!project || !descriptor || !descriptor.features?.includes("task_handoff")) return;
+        const client = await ControlClient.connect(project.stateDir, { role: "console", projectRoot: descriptor.cwd ?? project.root }, 3000);
+        try { await client.request({ t: "task_handoff_event", event: { projectId, remoteTask: task.id, state: event, ...(typeof task.sdlc?.pr === "number" ? { pr: task.sdlc.pr } : {}), ...(handoffHead(task) ? { head: handoffHead(task) } : {}) } }, 5000); }
+        finally { client.close(); }
+      } catch { /* the origin is down or refused: it reads the target's state on its next start */ }
+    })();
+  };
+  /**
+   * #604 ORIGIN side, on start: for every handoff this project made that has not reached a final state, ask the target (if it is running) where the task stands now, and record the answer.
+   * A target that is down is skipped: it is asked again on the next start. Best effort, in the background, never delaying readiness.
+   */
+  const handoffCatchUp = (): void => {
+    void (async () => {
+      const open = board.handoffOutbound().filter(row => !["approved", "cancelled"].includes(row.state));
+      const byProject = new Map<string, number[]>();
+      for (const row of open) byProject.set(row.remoteProject, [...(byProject.get(row.remoteProject) ?? []), row.remoteTask]);
+      for (const [remote, ids] of byProject) {
+        try {
+          const project = registeredProjects(hubHome()).find(candidate => candidate.id === remote);
+          const descriptor = project && readControl(project.stateDir);
+          if (!project || !descriptor?.features?.includes("task_handoff")) continue;
+          const client = await ControlClient.connect(project.stateDir, { role: "console", projectRoot: descriptor.cwd ?? project.root }, 3000);
+          try { const reply = await client.request({ t: "task_handoff_states", projectId, tasks: ids.slice(0, 200) }, 5000); if (reply.ok && Array.isArray(reply.states)) tasks.handoffCaughtUp(remote, reply.states); }
+          finally { client.close(); }
+        } catch { /* that target is down or refused: it is asked again on the next start */ }
+      }
+    })();
+  };
   const conductedMutations = new Set(["hub_task_propose", "hub_task_accept", "hub_task_decline", "hub_task_done", "hub_review", "hub_checkpoint", "hub_remember"]);
   const taskOp = async (...args: Parameters<typeof taskOpBody>): Promise<string> => {
     const [by, op, input] = args;
@@ -3377,7 +3520,7 @@ export async function startDaemon(opts: DaemonOptions) {
     }
     // Inside a PII turn the worker's words may carry the PII whatever they are attached to: a note would go to
     // shared memory and a new task could be routed to a cloud peer without matching any pattern.
-    if (piiTurn && (MEMORY_READ_TOOLS.has(op) || op === "hub_memory_correct" || op === "hub_remember" || op === "hub_task_propose" || op === "hub_sdlc_record" || op === "hub_plan_review" || op === "hub_sdlc_costs" || op === "hub_sdlc_metrics" || op.startsWith("hub_release_") || op === "hub_backlog_return_untouched" || op === "hub_backlog_plan" || op === "hub_backlog_undo" || op === "hub_backlog_pull" || op === "hub_backlog_split" || op === "release_new")) throw new PreEffectToolRefusal(`${op} is not available while working on a PII task: its text must not leave this machine`);
+    if (piiTurn && (MEMORY_READ_TOOLS.has(op) || op === "hub_memory_correct" || op === "hub_remember" || op === "hub_task_propose" || op === "hub_sdlc_record" || op === "hub_plan_review" || op === "hub_sdlc_costs" || op === "hub_sdlc_metrics" || op.startsWith("hub_release_") || op === "hub_backlog_return_untouched" || op === "hub_backlog_plan" || op === "hub_backlog_undo" || op === "hub_backlog_pull" || op === "hub_backlog_split" || op === "release_new" || op === "hub_study_draft" || op === "hub_study_revise")) throw new PreEffectToolRefusal(`${op} is not available while working on a PII task: its text must not leave this machine`);
     // The same words attached to an ordinary task would reach its reviewer, its owner, overlapping owners (a plan, the
     // completed-change notice) and shared memory. A turn that holds a PII task acts on ordinary tasks in a turn of its own.
     if (piiTurn && (op === "hub_task_done" || op === "hub_review" || (op === "hub_task_accept" && (a.plan != null || a.refs != null)))) {
@@ -3583,6 +3726,34 @@ export async function startDaemon(opts: DaemonOptions) {
       }
       case "hub_backlog_undo":
         return JSON.stringify(tasks.publicView(await tasks.undoPlanning(by, a.id, a.revision)));
+      // #601: the study tools only call Tasks. Draft and revise are the configured planner's (Tasks checks planningAuthority and the private turn); none of them approves.
+      // The arguments go to Tasks whole, never re-picked field by field: an unknown field is refused there (`studyFields`), not silently dropped here, and MCP schemas are not enforced on the way in.
+      case "hub_study_draft":
+        return JSON.stringify(await tasks.studyDraft(by, a));
+      case "hub_study_revise":
+        return JSON.stringify(await tasks.studyRevise(by, a));
+      case "hub_study_show":
+        return JSON.stringify(a.id === undefined ? tasks.studyList(by, a) : tasks.studyShow(by, a));
+      case "hub_study_handoff":
+        if (!a || typeof a !== "object" || Array.isArray(a) || Object.keys(a).some(key => key !== "handoff")) throw new PreEffectToolRefusal("study handoff takes only a handoff object");
+        return JSON.stringify(await tasks.studyHandoff(by, a.handoff));
+      // #604 (agent-hub#1030): a handoff to another project. Tasks checks who may send it and the PII gates; `handoffOut` resolves the target, asks its daemon and records the answer.
+      case "hub_task_handoff":
+        return JSON.stringify(await handoffOut(by, a));
+      case "hub_task_handoffs":
+        return JSON.stringify({ handoffs: tasks.handoffList() });
+      // #601: the person's own forms (a read is hub_study_show, which every connection may use). `by` is USER only on the console path; approval and close are never a tool.
+      case "study_draft":
+        return JSON.stringify(await tasks.studyDraft(by, a));
+      case "study_revise":
+        return JSON.stringify(await tasks.studyRevise(by, a));
+      case "study_approve":
+        return JSON.stringify(await tasks.studyApprove(by, a));
+      case "study_approve_protocol":
+        return JSON.stringify(await tasks.experimentApproveProtocol(by, a));
+      case "study_close":
+        return JSON.stringify(await tasks.studyClose(by, a));
+
       case "hub_task_propose": {
         const t = await tasks.propose(by, a);
         const repoHint = tasks.refsRepoHint(t); // #723: advice only; never part of the write
@@ -3790,7 +3961,7 @@ export async function startDaemon(opts: DaemonOptions) {
         const related = tasks.isPii(task) ? undefined : await taskMemoryResult(task, true).catch(() => ({ items: [], state: "unknown" as const }));
         const fresh = board.get(task.id);
         const relatedMemory = related && fresh?.updated === task.updated && !tasks.isPii(fresh) ? { state: related.state, count: related.items.length, ...("head" in related ? { head: related.head } : {}) } : { state: "unknown" };
-        return JSON.stringify({ ...detail, ...holderReasons(publicTask), relatedMemory, ...(codeSymbols ? { codeSymbols } : {}), ...(consoleDeliverables ? { deliverables: consoleDeliverables } : {}), symbolOverlaps: tasks.symbolOverlaps(task), flowRevision: task.history.length, reviews: board.reviews({ task: task.id }) }, null, 2);
+        return JSON.stringify({ ...detail, ...holderReasons(publicTask), relatedMemory, ...(publicTask.handoffOrigin ? { handoffOrigin: publicTask.handoffOrigin } : {}), ...(codeSymbols ? { codeSymbols } : {}), ...(consoleDeliverables ? { deliverables: consoleDeliverables } : {}), symbolOverlaps: tasks.symbolOverlaps(task), flowRevision: task.history.length, reviews: board.reviews({ task: task.id }) }, null, 2);
       }
       case "deliverables": {
         // #317 (`ahub deliverables`): the board's deliverable records, person only. `check` measures the files now; without it nothing is opened.
@@ -4134,6 +4305,7 @@ export async function startDaemon(opts: DaemonOptions) {
     if (tasks.autoPullsPending() !== 0) return false; // #785 P1c: an automatic backlog pull mid-admission would write after the commit's digest; its own line keeps the line above exact for scripts/seeds.json
     if (tasks.queuePressureRechecksPending() !== 0) return false;
     if (tasks.schedulesPending() !== 0) return false; // #512a: a scheduled run mid-admission (its PII screen is awaited) would write after the commit's digest; the hold fence refuses its write, and this waits for it to end
+    if (tasks.studiesPending() !== 0) return false;
     if (!recoveryPeerSnapshot) return true;
     const current = recoveryPeers();
     return recoveryPeerSnapshot.every((saved) => {
@@ -4265,6 +4437,7 @@ export async function startDaemon(opts: DaemonOptions) {
     const metadata = peer.recoveryMetadata?.() ?? {};
     const row: RestartPeerSnapshot = { id, state: peer.state, queueIds: bus.queueIds(id), ...(metadata.launch ? { launch: metadata.launch as Record<string, unknown> } : {}) };
     if (typeof metadata.threadId === "string") row.threadId = metadata.threadId;
+    if (typeof metadata.controlProtocol === "number" && Number.isSafeInteger(metadata.controlProtocol) && metadata.controlProtocol > 0) row.controlProtocol = metadata.controlProtocol;
     const sessionId = definitionOf(id)?.transport === "claude-channel" ? claudeSession(id).sessionId : metadata.sessionId;
     if (typeof sessionId === "string" && sessionId) row.sessionId = sessionId;
     return [id, row];
@@ -4283,7 +4456,12 @@ export async function startDaemon(opts: DaemonOptions) {
       return row;
     });
     const budgetState = budget.persistedPauseDigestRows().sort((a, b) => a.peer.localeCompare(b.peer));
-    return { queues, manualPaused: [...manualPaused].sort(), boardDigest: digest(boardState), budgetDigest: digest(budgetState), ...(allExternalInstanceIds().length ? { externalBindingsDigest: digest(allExternalInstanceIds().sort().map(id => [id, (externalRefusedRows.has(id) ? externalBindings.get(id) : retainedExternalBinding(id)) ?? null])) } : {}) };
+    const research = board.researchIntegrity();
+    const researchDispatch = board.researchDispatchIntegrity();
+    const researchExperiments = board.researchExperimentsIntegrity();
+    const researchComparisons = board.researchComparisonIntegrity();
+    const researchRunners = board.runnerReceiptIntegrity();
+    return { queues, manualPaused: [...manualPaused].sort(), boardDigest: digest(boardState), budgetDigest: digest(budgetState), ...(research.campaigns || research.units ? { research } : {}), ...(researchDispatch.activations || researchDispatch.handoffs ? { researchDispatch } : {}), ...(researchExperiments.protocols || researchExperiments.attempts || researchExperiments.events || researchExperiments.observations ? { researchExperiments } : {}), ...(researchComparisons.plans || researchComparisons.approvals || researchComparisons.contexts ? { researchComparisons } : {}), ...(researchRunners.count ? { researchRunners } : {}), ...(allExternalInstanceIds().length ? { externalBindingsDigest: digest(allExternalInstanceIds().sort().map(id => [id, (externalRefusedRows.has(id) ? externalBindings.get(id) : retainedExternalBinding(id)) ?? null])) } : {}) };
   };
   /** The integrity in the shape `expected` was recorded in: the current one unless only an older shape matches. */
   const integrityAs = (expected: unknown) => {
@@ -4373,7 +4551,7 @@ export async function startDaemon(opts: DaemonOptions) {
     ...(lifecycleConfig.mode !== "off" ? { lifecycleCriteria: true } : {}),
     // Every supported peer's membership, for launchers that run outside the console role (#484).
     roster: [...KNOWN_PEERS, ...OPTIONAL_PEERS].map(peer => ({ peer, on: onRoster(peer) })),
-    protocol: PROTOCOL,
+    protocol: PROTOCOL_MIN, protocolMin: PROTOCOL_MIN, protocolMax: PROTOCOL,
     features: DAEMON_FEATURES,
     configuredAdapters: configuredIds,
     ...(externalAdapters.length ? { externalAdapters } : {}),
@@ -4451,6 +4629,7 @@ export async function startDaemon(opts: DaemonOptions) {
   board.onChange = (task, entry) => {
     try { taskChanged?.(task, entry); }
     finally {
+      try { handoffNotifyOrigin(task, entry); } catch { /* never throw after a board write */ }
       try { settleSupersededDeliveries(); }
       finally { writeStatus(); } // committed task transitions refresh token-free attention even without bus activity
     }
@@ -4946,7 +5125,8 @@ export async function startDaemon(opts: DaemonOptions) {
     } else {
       log(`msg ${e.env.from} -> ${e.env.to?.join(",") ?? "*"} ${e.env.priority} hop=${e.env.hop}${e.dropped ? ` NOT DELIVERED(${e.dropped})` : ""}: ${e.env.body.slice(0, 200)}`);
       const env = e.env;
-      event({ type: "envelope", id: env.id, from: env.from, ...(env.to ? { to: env.to } : {}), priority: env.priority, hop: env.hop, ...(env.kind ? { kind: env.kind } : {}), ...(env.refs?.task ? { task: env.refs.task } : {}), ...(env.private ? {} : { bytes: Buffer.byteLength(env.body) }), ...(env.private ? { private: true } : {}), ...(e.dropped ? { dropped: e.dropped } : {}) });
+      const origin = takeOrigin(env.id) ?? adapterOrigin(env);
+      event({ type: "envelope", id: env.id, from: env.from, ...(origin ? { origin } : {}), ...(env.to ? { to: env.to } : {}), priority: env.priority, hop: env.hop, ...(env.kind ? { kind: env.kind } : {}), ...(env.refs?.task ? { task: env.refs.task } : {}), ...(env.private ? {} : { bytes: Buffer.byteLength(env.body) }), ...(env.private ? { private: true } : {}), ...(e.dropped ? { dropped: e.dropped } : {}) });
     }
     if (!stopping) writeStatus();
   });
@@ -6567,6 +6747,7 @@ export async function startDaemon(opts: DaemonOptions) {
     // Retired dictionaries never enter active publication; hub ids remain opaque keys.
     if (prefix === "routing.targets" || prefix === "routing.routes") return [];
     if (prefix === "routing.hub_routes") return [prefix];
+    if (prefix === "research.runners") return [prefix];
     if (value && typeof value === "object" && !Array.isArray(value)) return Object.entries(value).flatMap(([key, child]) => declaredPaths(child, prefix ? `${prefix}.${key}` : key));
     return prefix ? [prefix] : [];
   }
@@ -7386,6 +7567,7 @@ export async function startDaemon(opts: DaemonOptions) {
       ...(tasks.autoPullsPending() !== 0 ? ["an automatic backlog pull in flight"] : []),
       ...(tasks.queuePressureRechecksPending() !== 0 ? ["a queue-pressure assignment recheck in flight"] : []),
       ...(tasks.schedulesPending() !== 0 ? ["a scheduled task proposal in flight"] : []),
+      ...(tasks.studiesPending() !== 0 ? ["a research campaign admission in flight"] : []),
       ...(checkPending.size !== 0 ? ["isolated checks queued or running"] : []),
       ...(piCallsInFlight() !== 0 ? ["Pi tool calls in flight"] : []),
       ...(releasing ? ["ready tasks being released"] : []),
@@ -7427,6 +7609,44 @@ export async function startDaemon(opts: DaemonOptions) {
         && (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("task_planners"))) return "target lacks task_planners; configured or persisted planner authority cannot be restored by that daemon";
       if (board.releases().some(release => release.train && ["active", "frozen"].includes(release.state)) && (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("release_trains"))) return "target lacks release_trains; active or frozen train safeguards cannot be preserved";
       if ((extraIds.some(id => !!definitionOf(id)?.reviewer) || config.review.pool !== undefined) && (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("reviewer_pool"))) return "target lacks reviewer_pool; fresh review-only variants and pool policy cannot be preserved";
+      if (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_records")) {
+        if (board.hasResearchState()) return "target lacks research_records; persisted research campaign records require a supporting daemon";
+        let namesRunners = false;
+        try {
+          const texts = configSourceTexts(opts.cwd);
+          for (const name of CONFIG_FILES) {
+            if (texts[name] === null) continue;
+            const raw: unknown = JSON.parse(texts[name]!);
+            if (!raw || typeof raw !== "object" || Array.isArray(raw)) throw new Error("invalid configuration");
+            const research = (raw as Record<string, unknown>).research;
+            if (research !== undefined && (!research || typeof research !== "object" || Array.isArray(research))) throw new Error("invalid research configuration");
+            namesRunners ||= research !== undefined && Object.hasOwn(research, "runners");
+          }
+        } catch { return "research_records compatibility is unknown; configuration could not be read"; }
+        // Raw presence matters even for a disabled or ignored shared declaration: an older target parses the unknown key first.
+        if (namesRunners) return "target lacks research_records; the project's research.runners cannot be loaded by that daemon";
+      }
+      if (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_runners")) {
+        if (board.hasRunnerReceipts()) return "target lacks research_runners; persisted runner attestations require a supporting daemon";
+        try {
+          if (config.research.runners !== undefined || rawResearchRunnersPresent(opts.cwd)) return "target lacks research_runners; the project's research.runners requires a supporting daemon";
+        } catch { return "research_runners compatibility is unknown; configuration could not be read"; }
+      }
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_dispatch")) && board.hasResearchDispatchState())
+        return "target lacks research_dispatch; persisted campaign stop policies and handoffs require a supporting daemon";
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_dropped_offers")) && board.hasResearchDroppedOffers())
+        return "target lacks research_dropped_offers; persisted study re-offers require a supporting daemon";
+      // agent-hub#1036: an unsettled delivery that holds a distinct session's message is an envelope with `session: true` and a minted `<peer>~<id>` sender. A target without the feature
+      // would read that sender as a peer it does not know and queue a reply to it; this is its own label, so a daemon that has the older floors but not this one is still refused.
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("session_identity")) && bus.queueList(undefined, true).some(row => (row.originals as Envelope[]).some(env => env.session === true)))
+        return "target lacks session_identity; unsettled deliveries hold messages from distinct sessions that require a supporting daemon";
+      // #604: a handoff row in either direction (a task another project handed us, or one we handed over) needs a target that knows `task_handoff`; without it the origin is silently lost.
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("task_handoff")) && board.hasHandoffState())
+        return "target lacks task_handoff; stored handoffs require a supporting daemon";
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_experiments")) && board.hasResearchExperimentState())
+        return "target lacks research_experiments; persisted experiment records require a supporting daemon";
+      if ((!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("research_comparisons")) && board.hasResearchComparisonState())
+        return "target lacks research_comparisons; persisted comparison plans, approvals or protocol context require a supporting daemon";
       if (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("project_autonomy")) {
         let needsAutonomy = autonomyConfirmed.size > 0;
         try {
@@ -7450,8 +7670,29 @@ export async function startDaemon(opts: DaemonOptions) {
         } catch { return "action_policy_floor compatibility is unknown; configuration could not be read"; }
         if (namesFloor) return "target lacks action_policy_floor; the project's action_policy.floor cannot be loaded by that daemon";
       }
+      // #1017: an older target refuses a `cron` or `tz` key in a `schedules` entry ("cron not supported yet"), so the project would not load there. The raw files decide, as for the floor: a calendar
+      // entry in either file, even one an ordinary reload would reject, names the key; unreadable files are unknown, never absent.
+      if (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("calendar_schedules")) {
+        let namesCalendar = config.schedules.some(entry => entry.cron !== undefined || entry.tz !== undefined);
+        try {
+          const texts = configSourceTexts(opts.cwd);
+          namesCalendar ||= CONFIG_FILES.some(name => {
+            const list = texts[name] === null ? undefined : (JSON.parse(texts[name]!) as Record<string, unknown>).schedules;
+            return Array.isArray(list) && list.some(entry => !!entry && typeof entry === "object" && (Object.hasOwn(entry, "cron") || Object.hasOwn(entry, "tz")));
+          });
+        } catch { return "calendar_schedules compatibility is unknown; configuration could not be read"; }
+        if (namesCalendar) return "target lacks calendar_schedules; the project's schedules use cron or tz, which that daemon cannot load";
+      }
+      if (!Array.isArray(msg.targetFeatures) || !msg.targetFeatures.includes("schema_guard"))
+        return "target lacks schema_guard; controlled recovery requires a guard-aware runtime";
       return undefined;
     };
+    // Preserve lifecycle fences before capability probes; the prepare branch rechecks them after resume capture awaits.
+    if (msg.op === "prepare") {
+      if (rosterChanging) return recoveryError("roster change is in progress; wait for ahub status before preparing recovery");
+      if (requestedPeerStops.size) return recoveryError("peer stop is in progress; wait for ahub status before preparing recovery");
+      if (recoveryCommitted) return recoveryError("recovery commit is already in progress");
+    }
     const checkTarget = msg.op === "prepare" || msg.op === "commit" || msg.op === "inspect" && msg.targetFeatures !== undefined;
     if (checkTarget) {
       const error = incompatibleRecoveryTarget(); if (error) return recoveryError(error);
@@ -7719,7 +7960,8 @@ export async function startDaemon(opts: DaemonOptions) {
     if (!c.authed) {
       if (msg.t !== "hello" || msg.token !== token) return sock.close(4401, "bad token");
       if (typeof msg.peer === "string" && msg.peer.includes("@")) return sock.close(4403, "remote identities have no control connection");
-      if (msg.v !== PROTOCOL) {
+      const offered = protocolWindow(msg.v, msg.vMin), selectedProtocol = offered && negotiateProtocol({ v: PROTOCOL, vMin: PROTOCOL_MIN }, offered);
+      if (selectedProtocol === undefined) {
         // An outdated plugin would drop every digest without a trace. Refuse it loudly instead.
         log(`refused ${msg.role} ${msg.peer ?? ""}: wire version ${msg.v ?? 1}, hub speaks ${PROTOCOL} (claude plugin update agent-hub@agent-hub)`);
         return sock.close(4426, `wire version mismatch: hub speaks ${PROTOCOL}; update the agent-hub plugin`);
@@ -7730,14 +7972,25 @@ export async function startDaemon(opts: DaemonOptions) {
       try { refreshConductorPolicy(); } catch { return sock.close(4403, "invalid conductor role/feed configuration"); }
       if (stopping && msg.role !== "console") return sock.close(1013, "hub is stopping");
       c.authed = true;
+      c.protocol = selectedProtocol;
       c.role = msg.role;
       if (c.role === "console") consoles.add(sock);
+      else if (c.role === "tools" && sessionHello(msg)) {
+        // agent-hub#1036: a connection that names a peer id but is not that peer's current session gets an id of its own instead of becoming the peer. It can send and read; it holds none of the
+        // peer's board roles, is never a bus peer (no delivery, no push) and replaces nobody. Nothing is refused: a hello that worked before still works, as a session.
+        const named = String(msg.peer ?? "");
+        if (!PEER_ID.test(named) || named === USER || named === "hub") return sock.close(4403, "peer id is reserved or malformed");
+        c.peer = mintSessionId(named, msg.session, id => sessionIds.has(id) || bus.peers.has(id), () => crypto.randomUUID().slice(0, 6));
+        c.session = true; sessionIds.add(c.peer);
+        if (typeof msg.launchId === "string" && msg.launchId.length <= 128) c.launchId = msg.launchId;
+      }
       else if (c.role === "tools") {
         // Acts for a peer the hub manages itself (kimi, codex): may send and use the board as that peer, is never a delivery target.
         c.peer = String(msg.peer ?? "");
         if (!PEER_ID.test(c.peer) || c.peer === USER || c.peer === "hub") return sock.close(4403, "peer id is reserved or malformed");
         if (externalRefusedRows.has(c.peer)) return sock.close(4403, "instance kind is unavailable in the startup catalog");
         if (!membershipAllowed(c.peer)) return sock.close(4000, "peer is off roster; retry when a person adds it");
+        if (typeof msg.launchId === "string" && msg.launchId.length <= 128) c.launchId = msg.launchId;
         if (definitionOf(c.peer)?.start.daemonPort === "claude-implementer") {
           const owner = bus.peers.get(c.peer), launch = owner?.ownedTools?.(msg.launchId), root = ownerWriteRoot(c.peer);
           if (!owner || !launch || launch.signal.aborted || !root || root === opts.cwd || typeof msg.launchId !== "string") return sock.close(4403, "this process is not the current Claude child");
@@ -7804,7 +8057,9 @@ export async function startDaemon(opts: DaemonOptions) {
           if (important) ws.notifyQueued({ queued: bus.queued(c.peer!), important });
         });
       }
-      return void reply({ t: "welcome", projectId, instanceId, cwd: opts.cwd, protocol: PROTOCOL, features: DAEMON_FEATURES });
+      // #604 PR B: one attach event per admitted peer or tools hello, with the same fields an envelope's `origin` carries. A refused hello returned above and writes none.
+      if (c.role === "peer" || c.role === "tools") event(attachEvent({ role: c.role, peer: c.peer, launchId: c.launchId, ...(c.session ? { session: true } : {}) }));
+      return void reply({ t: "welcome", projectId, instanceId, cwd: opts.cwd, protocol: c.protocol, protocolMin: PROTOCOL_MIN, protocolMax: PROTOCOL, features: DAEMON_FEATURES, ...(c.session ? { session: c.peer } : {}) });
     }
     const committedSourceProof = c.role === "tools" && msg.t === "recovery_authority" && msg.op === "verify" && recoveryCommitted
       && msg.binding?.operationId === recoveryOperationId && msg.binding?.sourceInstanceId === instanceId;
@@ -7812,6 +8067,11 @@ export async function startDaemon(opts: DaemonOptions) {
     if (c.role !== "console" && c.peer && !membershipAllowed(c.peer)) return void reply({ ok: false, error: `${c.peer} is off roster` });
     if (c.externalOwner && !currentExternalTools(c)) return void reply({ ok: false, error: "external instance launch or execution root changed" });
     if (c.toolsOwner && !currentImplementerTools(c)) return void reply({ ok: false, error: "Claude child generation or execution root changed" });
+    if (c.session) {
+      // agent-hub#1036: a distinct session sends and reads, and nothing else. The gate runs before any case, so a new control message is refused to a session unless it is added to the list.
+      const refusal = sessionGate(c.peer!, msg);
+      if (refusal) return void reply({ t: msg.t, ok: false, error: refusal });
+    }
     switch (msg.t) {
       case "request_cancel": {
         if (!Number.isSafeInteger(msg.request) || msg.request < 1) return void reply({ t: "request_cancel", ok: false, error: "invalid request id" });
@@ -8068,11 +8328,15 @@ export async function startDaemon(opts: DaemonOptions) {
             const plan = msg.plan as UpgradePlan;
             const { fingerprint, ...body } = plan ?? {} as UpgradePlan;
             if (!plan || !["restart", "upgrade"].includes(plan.kind) || plan.projects.some(row => row.terminals.length > 0) || planFingerprint(body) !== fingerprint || fingerprint !== raw.planFingerprint || plan.kind !== raw.kind || existsSync(operationPath(raw.operationId))
-              || raw.sourceInstanceId !== instanceId || recoveryActive() || recoveryLock() || !plan.projects.some(row => row.project.id === projectId && row.project.root === opts.cwd && row.project.stateDir === opts.stateDir && row.source.instanceId === instanceId)) throw new RecoveryAuthorityError();
+              || raw.sourceInstanceId !== instanceId || recoveryActive() || recoveryLock(undefined, opts.cwd) || !plan.projects.some(row => row.project.id === projectId && row.project.root === opts.cwd && row.project.stateDir === opts.stateDir && row.source.instanceId === instanceId)) throw new RecoveryAuthorityError();
+            const { recoveryScope } = await import("../cli/upgrade.ts"); current();
+            const scope = recoveryScope(plan);
+            if (scope !== undefined && scope !== opts.cwd) throw new RecoveryAuthorityError();
             if (plan.kind === "restart" && plan.projects.length !== 1) throw new RecoveryAuthorityError();
             if (plan.kind === "upgrade") {
-              // Shared package/plugin effects cover every running or uncertain registered project, not a caller-selected subset.
+              // Scoped upgrades bind the source's one registered project; retained shared upgrades still cover every affected project.
               const affected = registeredProjects().filter(project => {
+                if (scope !== undefined) return project.root === scope;
                 const control = readControl(project.stateDir);
                 return control ? processLiveness(control.pid, control.pidSignature) !== "gone" : project.pid !== null && processLiveness(project.pid, project.pidSignature) !== "gone";
               });
@@ -8085,23 +8349,26 @@ export async function startDaemon(opts: DaemonOptions) {
               if (!Array.isArray(selected.reconnectOnly) || new Set(selected.reconnectOnly).size !== selected.reconnectOnly.length) throw new RecoveryAuthorityError();
               for (const id of selected.reconnectOnly) {
                 const row = selected.source.peers.find(peer => peer.id === id), owner = bus.peers.get(id), definition = definitionOf(id);
-                if (!row || selected.freshStart?.includes(id) || Object.keys(row).some(key => !["id", "state", "observation"].includes(key)) || !row.observation
+                if (!row || selected.freshStart?.includes(id) || Object.keys(row).some(key => !["id", "state", "observation", "controlProtocol"].includes(key)) || !row.observation
                   || Object.keys(row.observation).some(key => !["kind", "attached", "sessionDigest"].includes(key))
                   || !isWsPeer(owner) || definition?.transport !== "claude-channel" || !owner.attached || owner.managedLaunchId || claudeLaunchProof.verified(id)
+                  || !protocolWindow(row.controlProtocol) || owner.recoveryMetadata().controlProtocol !== row.controlProtocol
                   || row.observation.kind !== definition.kind || row.observation.attached !== owner.attached || row.observation.sessionDigest !== recoverySessionDigest(claudeSession(id).sessionId)) throw new RecoveryAuthorityError();
                 const generation = owner.sessionGeneration;
                 const launcher = await launcherOf(id, opts.cwd, { stateDir: opts.stateDir, instanceId }); current();
-                if (launcher && launcher.state !== "gone" || bus.peers.get(id) !== owner || owner.sessionGeneration !== generation || !owner.attached || owner.managedLaunchId || claudeLaunchProof.verified(id)) throw new RecoveryAuthorityError();
+                if (launcher && launcher.state !== "gone" || bus.peers.get(id) !== owner || owner.sessionGeneration !== generation || !owner.attached || owner.managedLaunchId || claudeLaunchProof.verified(id)
+                  || owner.recoveryMetadata().controlProtocol !== row.controlProtocol) throw new RecoveryAuthorityError();
                 reconnectCurrent.push(() => bus.peers.get(id) === owner && owner.sessionGeneration === generation && owner.attached && !owner.managedLaunchId && !claudeLaunchProof.verified(id)
+                  && owner.recoveryMetadata().controlProtocol === row.controlProtocol
                   && definitionOf(id)?.kind === row.observation!.kind && owner.attached === row.observation!.attached && recoverySessionDigest(claudeSession(id).sessionId) === row.observation!.sessionDigest);
                 reconnect.set(id, row.observation);
               }
             }
             if (reconnectCurrent.some(check => !check())) throw new RecoveryAuthorityError();
-            const normalize = (row: { id: string; state: string; observation?: { kind: string; attached: boolean; sessionDigest?: string }; threadId?: string; sessionId?: string; sessionFile?: string; args?: Record<string, string>; launch?: Record<string, unknown> }) => {
+            const normalize = (row: { id: string; state: string; controlProtocol?: number; observation?: { kind: string; attached: boolean; sessionDigest?: string }; threadId?: string; sessionId?: string; sessionFile?: string; args?: Record<string, string>; launch?: Record<string, unknown> }) => {
               const launch = row.args ?? row.launch ?? {};
               const observed = reconnect.get(row.id);
-              return { id: row.id, active: row.state !== "offline", threadId: row.threadId, sessionId: observed ? undefined : row.sessionId,
+              return { id: row.id, active: row.state !== "offline", controlProtocol: row.controlProtocol, threadId: row.threadId, sessionId: observed ? undefined : row.sessionId,
                 ...(observed ? { sessionDigest: row.observation?.sessionDigest ?? recoverySessionDigest(row.sessionId) } : {}),
                 ...(observed ? { kind: row.observation?.kind ?? definitionOf(row.id)?.kind, attached: row.observation?.attached ?? (isWsPeer(bus.peers.get(row.id)) && (bus.peers.get(row.id) as import("../adapters/channel-peer.ts").WsPeer).attached) } : {}),
                 sessionFile: row.sessionFile ?? (typeof launch.sessionFile === "string" ? launch.sessionFile : undefined),
@@ -8115,9 +8382,11 @@ export async function startDaemon(opts: DaemonOptions) {
             if (raw.kind === "restart") {
               if (!sourcePackageDigest || raw.targetDigest !== sourcePackageDigest) throw new RecoveryAuthorityError();
             } else {
-              const staged = join(hubHome(), "releases", plan.version);
-              if (raw.targetRoot !== join(staged, "node_modules/@staix/agent-hub")) throw new RecoveryAuthorityError();
-              const verified = JSON.parse(readFileSync(join(staged, "verified.json"), "utf8"));
+              const staged = join(scope === undefined ? hubHome() : realPath(hubHome()), "releases", plan.version);
+              const expectedTarget = scope === undefined ? join(staged, "node_modules/@staix/agent-hub") : join(staged, `source-${raw.targetDigest}`);
+              if (raw.targetRoot !== expectedTarget) throw new RecoveryAuthorityError();
+              // A new/expired grant still needs registry proof; verification of an already issued grant below uses only retained target bytes.
+              const verified = JSON.parse(readFileSync(join(scope === undefined ? staged : join(staged, "registry"), "verified.json"), "utf8"));
               if (verified.digest !== raw.targetDigest || verified.integrity !== plan.integrity) throw new RecoveryAuthorityError();
             }
             if (!sourcePackageDigest || plan.sourceDigest !== sourcePackageDigest && (plan.kind !== "upgrade" || plan.sourceDigest !== raw.targetDigest)) throw new RecoveryAuthorityError();
@@ -8184,6 +8453,8 @@ export async function startDaemon(opts: DaemonOptions) {
         }
         // Built first: limits count the audience it really has (a reply goes to the parent's sender).
         let env = newEnvelope(c.peer ?? USER, body, { priority, ...(to ? { to } : {}), ...(inReplyTo ? { inReplyTo } : {}) });
+        if (c.session) env = { ...env, session: true };   // agent-hub#1036: sent by a distinct session; `from` is its minted id, and the header a recipient reads shows it as is
+        rememberOrigin(env.id, originOf({ role: c.role, peer: c.peer, launchId: c.launchId, ...(c.session ? { session: true } : {}) })); // the connection that sent it, for events.jsonl (never a path or text)
         try { bus.checkRemotePublication(env); }
         catch (error) { return void reply({ t: "sent", ok: false, error: error instanceof Error && /^(?:room|remote) /.test(error.message) ? sanitize(error.message).slice(0, 8192) : "message admission unavailable" }); }
         const offRoster = bus.offRosterRecipients(env);
@@ -8802,6 +9073,8 @@ export async function startDaemon(opts: DaemonOptions) {
           try { const current = publicReadCurrent(sock); current(); const release = publicReleaseShow(msg.args ?? {}); current(); return void reply({ t: "task", ok: true, text: JSON.stringify(release) }); }
           catch { return void reply({ t: "task", ok: false, error: "public release read unavailable" }); }
         }
+        // #601: ahub study is the person's. A tools connection (any agent) may use the hub_study_* tools, never these op names, whatever Tasks would answer.
+        if (String(msg.op).startsWith("study_") && c.role !== "console") return void reply({ t: "task", ok: false, error: "ahub study draft, revise, approve and close are console commands; agents use the hub_study_* tools" });
         if (c.toolsOwner && !IMPLEMENTER_HUB_TOOL_NAMES.includes(String(msg.op))) return void reply({ t: "task", ok: false, error: "tool is not available to this Claude implementer child" });
         if (recoveryActive() && recoveryPhase !== "preparing" && String(msg.op) !== "hub_checkpoint") return void reply({ t: "task", ok: false, error: "recovery is holding mutations" });
         if (c.role === "tools" && c.peer && Object.hasOwn(CONDUCTOR_TASK_OPS, String(msg.op))) {
@@ -8973,6 +9246,32 @@ export async function startDaemon(opts: DaemonOptions) {
         // Let the acknowledgement flush before closing the control listener.
         setTimeout(() => void stop().catch((error) => log(`shutdown incomplete: ${(error as Error).message}`)), 0);
         return;
+      // #604 (agent-hub#1030): another project's daemon hands us a task, or reports a state change of one we handed it. Console role only (the same token and the same machine user as every
+      // control message): an agent connection never reaches these. The origin peer named in the request is a record, never an identity here, and nothing in it grants a role.
+      case "task_handoff": {
+        if (c.role !== "console") return void reply({ t: "task_handoff", ok: false, error: "a handoff comes from another project's hub, not from an agent connection" });
+        if (recoveryActive()) return void reply({ t: "task_handoff", ok: false, error: "recovery is holding mutations; hand the task off again after it ends" });
+        // Counted like a task op: a recovery commit waits for it (`recoveryReady` reads `taskOpsInFlight`), and the fence is asked again right before the board write, after the PII screen.
+        taskOpsInFlight++;
+        const recoveryFence = (): string | undefined => recoveryActive() || stopping ? "recovery or shutdown is holding mutations; hand the task off again after it ends" : undefined;
+        void tasks.handoffIn(msg.origin ?? {}, msg.request, recoveryFence).then(
+          result => reply({ t: "task_handoff", ok: true, ...result }),
+          error => reply({ t: "task_handoff", ok: false, error: error instanceof HandoffRefusal ? error.message : "the handoff was refused" })).finally(() => { taskOpsInFlight--; });
+        return;
+      }
+      case "task_handoff_event": {
+        if (c.role !== "console") return void reply({ t: "task_handoff_event", ok: false, error: "a handoff event comes from another project's hub, not from an agent connection" });
+        try { reply({ t: "task_handoff_event", ok: true, ...tasks.handoffEvent(msg.event) }); }
+        catch (error) { reply({ t: "task_handoff_event", ok: false, error: error instanceof HandoffRefusal ? error.message : "the handoff event was refused" }); }
+        return;
+      }
+      case "task_handoff_states": {
+        // #604: the origin asks, after a restart, where the tasks it handed us stand now. Console role only, and only for tasks whose recorded origin is the asking project.
+        if (c.role !== "console") return void reply({ t: "task_handoff_states", ok: false, error: "a handoff state request comes from another project's hub, not from an agent connection" });
+        try { reply({ t: "task_handoff_states", ok: true, states: tasks.handoffStatesFor(msg.projectId, msg.tasks) }); }
+        catch (error) { reply({ t: "task_handoff_states", ok: false, error: error instanceof HandoffRefusal ? error.message : "the handoff state request was refused" }); }
+        return;
+      }
       default:
         // A newer CLI talking to an older hub must get an answer, not wait forever.
         if (msg.rid !== undefined) reply({ t: String(msg.t), ok: false, error: `this hub does not know "${msg.t}" (restart it: ahub kill && ahub up)` });
@@ -8991,6 +9290,7 @@ export async function startDaemon(opts: DaemonOptions) {
     stopping = true;
     roomCommands.stop();
     tasks.stopQueuePressureRechecks();
+    const studiesStopped = tasks.stopStudies();
     let implementerShellCleanupFailed = false;
     const implementerShellStopped = implementerShellStart?.then(wrapper => wrapper.close()).catch(() => { implementerShellCleanupFailed = true; log("host shell cleanup failed during shutdown"); });
     terminalLaunches.clear();
@@ -9013,11 +9313,11 @@ export async function startDaemon(opts: DaemonOptions) {
     for (const pending of checkpointWaits.values()) pending.done(undefined);
     for (const p of permissions.values()) p.done(undefined);
     try {
-      // A schedules pass may be inside its PII screen: `stopping` is set, so its fence refuses the write when the screen returns. Wait for it (bounded, the screen has
-      // its own timeout) before the peers stop, the board closes and the store closes, so no pass runs against a closed database.
-      if (schedulesPass || manualSchedulePasses.size) {
+      // A schedule or study admission may be inside its PII screen: `stopping` is set, so its fence refuses the write when the screen returns. Wait for it (bounded, the screen has
+      // its own timeout) before the peers stop, the board closes and the store closes, so no admission runs against a closed database.
+      if (schedulesPass || manualSchedulePasses.size || tasks.studiesPending() !== 0) {
         let timer: ReturnType<typeof setTimeout> | undefined;
-        await Promise.race([Promise.all([schedulesPass?.catch(() => undefined), ...manualSchedulePasses]), new Promise<void>(resolve => { timer = setTimeout(resolve, 5_000); timer.unref?.(); })]);
+        await Promise.race([Promise.all([schedulesPass?.catch(() => undefined), ...manualSchedulePasses, studiesStopped]), new Promise<void>(resolve => { timer = setTimeout(resolve, 5_000); timer.unref?.(); })]);
         clearTimeout(timer);
       }
       await Promise.all([terminalReadsStopped, codeFactsStopped, implementerShellStopped]);
@@ -9087,6 +9387,7 @@ export async function startDaemon(opts: DaemonOptions) {
   bus.onQueues(); // initial restored holds are reported only after the listener is bound; writes status.json too
   log(`${RUN_START}${process.pid} control=127.0.0.1:${server.port} cwd=${opts.cwd}`);
   ready = true;
+  handoffCatchUp();
   tasks.scheduleQueuePressureRecheck();
   for (const name of opts.configWatch === false ? [] : CONFIG_SOURCE_FILES) {
     const path = join(opts.cwd, ".agenthub", name);