@bli-cockpit/mcp 0.1.121 → 0.1.125

package/README.md

@@ -109,7 +109,7 @@ about what an agent can reach.
 
 <!-- BEGIN GENERATED verb census — `npm run mcp:readme` -->
 
-**171 of 176 Tower verbs have an MCP twin.**
+**188 of 193 Tower verbs have an MCP twin.**
 Each tool goes through the SAME door its CLI verb calls, with the same
 collector device token — never a second route and never a service-role
 reader. `src/verb-census.test.ts` fails when a verb is in none of the
@@ -206,6 +206,22 @@ three tables below.
 | `cockpit mail send` | `mail_send` | `POST /api/mail/send` |
 | `cockpit mail sync` | `mail_sync` | `POST /api/mail/accounts/[id]/sync` |
 | `cockpit mail sync-status` | `mail_sync_status` | `GET /api/mail/sync-status` |
+| `cockpit manager add` | `manager_add` | `POST /api/manager/clients/[clientId]/actions (or /timeline)` |
+| `cockpit manager approve` | `manager_approve` | `POST /api/manager/actions/[id]/review (or /timeline/[id]/review)` |
+| `cockpit manager archive` | `manager_archive` | `POST /api/manager/actions/[id]/archive (or /timeline/[id]/archive)` |
+| `cockpit manager client` | `manager_client` | `GET /api/manager/clients/[clientId]` |
+| `cockpit manager clients` | `manager_clients` | `GET /api/manager/clients` |
+| `cockpit manager dismiss` | `manager_dismiss` | `POST /api/manager/actions/[id]/review (or /timeline/[id]/review)` |
+| `cockpit manager done` | `manager_done` | `POST /api/manager/actions/[id]/status` |
+| `cockpit manager edit` | `manager_edit` | `PATCH /api/manager/actions/[id] (or /timeline/[id])` |
+| `cockpit manager link` | `manager_link` | `POST /api/manager/actions/[id]/link` |
+| `cockpit manager list` | `manager_list` | `GET /api/manager/clients/[clientId]/actions (or /timeline)` |
+| `cockpit manager proposal` | `manager_proposal` | `POST /api/manager/actions/[id]/proposal (or /timeline/[id]/proposal)` |
+| `cockpit manager queue` | `manager_queue` | `GET /api/manager/clients/[clientId]/review-queue` |
+| `cockpit manager reopen` | `manager_reopen` | `POST /api/manager/actions/[id]/status` |
+| `cockpit manager run` | `manager_run` | `GET /api/manager/runs/[runId]` |
+| `cockpit manager show` | `manager_show` | `GET /api/manager/actions/[id] (or /timeline/[id])` |
+| `cockpit manager unarchive` | `manager_unarchive` | `POST /api/manager/actions/[id]/unarchive (or /timeline/[id]/unarchive)` |
 | `cockpit memory chain` | `memory_chain` | `GET /api/memory/chain` |
 | `cockpit memory consolidation` | `memory_consolidation` | `GET /api/memory/consolidation` |
 | `cockpit memory forget` | `forget_memory` (`bli-memory`) | `POST /api/memory/forget` |
@@ -277,6 +293,7 @@ three tables below.
 | `cockpit team device doctor` | `team_device_doctor` | `POST /api/ambient/devices/[deviceId]/doctor` |
 | `cockpit team device list` | `team_device_list` | `GET /api/team/devices` |
 | `cockpit team device revoke` | `team_device_revoke` | `POST /api/ambient/devices/[deviceId]/revoke` |
+| `cockpit team device scope` | `team_device_scope` | `POST /api/ambient/devices/[deviceId]/scope` |
 | `cockpit team github-login add` | `team_github_login_add` | `POST /api/team/members/[userId]/github-logins` |
 | `cockpit team github-login list` | `team_github_login_list` | `GET /api/team/members/[userId]/github-logins` |
 | `cockpit team github-login remove` | `team_github_login_remove` | `DELETE /api/team/members/[userId]/github-logins` |
@@ -304,7 +321,7 @@ A claim about the verb's nature, not a backlog.
 | `cockpit brief edit` | opens the person's own $EDITOR on this machine and files what they changed; an agent has no editor to open (commands/editor.ts) |
 | `cockpit cal add-ical` | carries a calendar's SECRET iCal ADDRESS — a permanent, unauthenticated, read-anything-on-that-calendar URL. The CLI reads it from STDIN precisely so it never lands in an argument list; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so attaching a calendar stays a thing a person does at a terminal (BLI-3709) |
 | `cockpit mail add-imap` | carries a Google app password. The CLI reads it from STDIN precisely so it never lands in an argument list; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so attaching a mailbox stays a thing a person does at a terminal (BLI-3708) |
-| `cockpit notes connect` | carries a notetaker credential: an API key (Fellow or Circleback), or for Wispr Flow a browser sign-in whose refresh token the CLI hands straight to Tower. The CLI reads a key from STDIN precisely so it never lands in an argument list, and a browser sign-in has no tool-call equivalent at all; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so connecting a notetaker stays a thing a person does at a terminal or in Tower's Settings (BLI-4383, BLI-12926) |
+| `cockpit notes connect` | an API-key connect (Fellow or Circleback) carries a notetaker credential. The CLI reads the key from STDIN precisely so it never lands in an argument list, and an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so connecting a key stays a thing a person does at a terminal or in Tower's Settings (BLI-4383). Wispr Flow carries no credential through a call: its sign-in is a link the person clicks, started and watched over the same doors by the MCP tools notes_connect_wisprflow and notes_connect_status (BLI-17973) |
 | `cockpit notes webhook` | carries the VENDOR's webhook signing secret, the key an HMAC over every delivered payload is checked against. Exactly the ruling `notes connect` already carries, for the same credential shape: the CLI reads it from STDIN so it never lands in an argument list, and an MCP tool argument travels through a model's context window and whatever transcript store sits behind it (BLI-4394) |
 
 <!-- END GENERATED verb census -->

package/dist/device-twins.d.ts

@@ -0,0 +1,10 @@
+/**
+ * The device-scope twin in the verb census (BLI-17978).
+ *
+ * `cockpit team device scope <device> notetaker_bot|none` and `team_device_scope` both ride
+ * `POST /api/ambient/devices/[deviceId]/scope`, the one door that gives the meeting bot's device the
+ * scope Tower's claim door requires (or takes it away). Split out of `verb-census.ts`, which sits at
+ * the file-size ratchet.
+ */
+import type { McpTwin } from "./verb-census.js";
+export declare const DEVICE_TWINS: Record<string, McpTwin>;

package/dist/device-twins.js

@@ -0,0 +1,11 @@
+/**
+ * The device-scope twin in the verb census (BLI-17978).
+ *
+ * `cockpit team device scope <device> notetaker_bot|none` and `team_device_scope` both ride
+ * `POST /api/ambient/devices/[deviceId]/scope`, the one door that gives the meeting bot's device the
+ * scope Tower's claim door requires (or takes it away). Split out of `verb-census.ts`, which sits at
+ * the file-size ratchet.
+ */
+export const DEVICE_TWINS = {
+    "team device-scope": { tool: "team_device_scope", door: "POST /api/ambient/devices/[deviceId]/scope" },
+};

package/dist/manager-tools.d.ts

@@ -0,0 +1,7 @@
+import { type AgentDoorResponse } from "./agent-door.js";
+import { type ToolDeps } from "./tool-result.js";
+/** The Manager's refusal in the person's words: label, finer label when named, then the sentence. */
+export declare function managerFailureText(tool: string, response: AgentDoorResponse): string;
+export declare function registerManagerTools(server: {
+    registerTool: (...args: never[]) => unknown;
+}, deps: ToolDeps): void;

package/dist/manager-tools.js

@@ -0,0 +1,363 @@
+/**
+ * Tower Manager over MCP (BLI-5006): the sixteen twins of `cockpit manager`.
+ *
+ * Plain words: the Manager keeps, per client, a timeline (meetings, documents,
+ * milestones, internal events) and a list of actions (to-dos with owners and a
+ * due date). The AI proposes items and changes; a person reviews them. These tools
+ * let an agent read both and do what a person does on the page: add, edit, tick off,
+ * link, approve or dismiss the AI's suggestion, archive. They ride the SAME
+ * `/api/manager/**` door the CLI calls, on this machine's device token; the door
+ * hands the Manager's own answer back unchanged.
+ *
+ * Every write on an existing item carries `version`, the item's `updatedAt` as last
+ * read, sent as `expectedUpdatedAt`: a stale one is refused (409 invalid_state,
+ * `refusal item_changed`) rather than written over someone else's change.
+ *
+ * Deciding the AI's suggestion (approve, dismiss, proposal) is not an outward act, so
+ * none of these needs `confirm`: it records the person whose key this is as the
+ * reviewer and takes the item out of the review queue unless a suggested change still waits on it. Nothing here reaches anyone.
+ *
+ * Refusals come in two shapes and are relayed in the person's own words
+ * (`managerFailureText`): the Manager's `{ error: <sentence>, reason, refusal? }` and
+ * the door's own `{ code, reason, message }`. Only these tools read the first shape;
+ * every other family's refusal text is untouched.
+ */
+import { randomUUID } from "node:crypto";
+import { z } from "zod";
+import { callAgentDoor } from "./agent-door.js";
+import { dataResult, doorFailureText, doorReason, errorResult, queryString, registrarFor, withSession, } from "./tool-result.js";
+const KINDS = ["meeting", "document", "milestone", "internal"];
+const ACTION_STATUSES = ["open", "done", "dropped", "closed", "all"];
+const DEFAULT_LIMIT = 50;
+const MAX_LIMIT = 200;
+const CLIENT_ID = z.string().uuid().describe("The client's id, from manager_clients.");
+const ITEM_ID = z.string().uuid().describe("The item's id, from manager_list, manager_queue or manager_show.");
+const TIMELINE = z.boolean().optional()
+    .describe("true when the item is a timeline item (a meeting, document, milestone or internal event). Omit or false for an action (a to-do).");
+const VERSION = z.string().min(1)
+    .describe("The item's updatedAt exactly as manager_list or manager_show returned it. If anyone changed the item since, the write is refused (item_changed): read it again and retry.");
+const DATE = z.string().regex(/^\d{4}-\d{2}-\d{2}$/, "a date written YYYY-MM-DD");
+const OWNERS = z.array(z.string().min(1).max(120)).max(20).optional()
+    .describe("Actions only: the owners, one person per entry. Leaving this out leaves the owners alone; an empty list clears them; a list replaces them.");
+const OWNER_NAME = z.string().max(120).nullable().optional()
+    .describe("Timeline items only: the one owner's name. Leaving this out leaves the owner alone; null (or an empty string) clears it.");
+/** The fields only one table has; naming one on the other table is refused here, never dropped silently. */
+const ACTION_ONLY = ["owners", "due", "meeting_id", "status", "owner", "no_owner", "overdue"];
+const TIMELINE_ONLY = ["kind", "date", "owner_name"];
+/** The Manager's refusal in the person's words: label, finer label when named, then the sentence. */
+export function managerFailureText(tool, response) {
+    if (response.transportError)
+        return doorFailureText(tool, response);
+    const body = response.body;
+    const str = (key) => (typeof body[key] === "string" && body[key] !== "" ? body[key] : null);
+    const sentence = str("message") ?? (typeof body["reason"] === "string" ? str("error") : null) ?? `Tower answered ${response.status}.`;
+    const refusal = str("refusal");
+    const reason = doorReason(body);
+    // A change with no answer may or may not have landed (an add says to retry with its request_id instead).
+    const unknown = (reason === "manager_timeout" || reason === "manager_unreachable" || response.status >= 500) && tool !== "manager_add" && !READ_TOOLS_HERE.has(tool);
+    const next = NEXT_STEP[refusal ?? ""] ?? (reason === "not_found" ? NEXT_STEP["not_found"] : unknown ? NEXT_STEP["no_answer"] : undefined);
+    return `Tower refused ${tool} (${doorReason(body)}${refusal ? `, refusal ${refusal}` : ""}): ${sentence}${next ? ` ${next}` : ""}`;
+}
+/** What to do next after a refusal, in tool terms, so every refusal ends with a way forward (the CLI says the same). */
+const READ_TOOLS_HERE = new Set(["manager_clients", "manager_client", "manager_run", "manager_list", "manager_queue", "manager_show"]);
+const NEXT_STEP = {
+    no_answer: "It may or may not have been saved: read the item with manager_show, and use the version it gives if you try again.",
+    not_found: "Check the id with manager_list; a timeline item needs timeline true, and done, reopen and link are for actions only.",
+    item_changed: "Read the item again with manager_show for its current version, then retry.",
+    item_archived: "Bring it back first with manager_unarchive (manager_list with archived true gives its version).",
+    item_dismissed: "Restore it first with manager_approve and restore true (manager_list with review dismissed gives its version).",
+    person_item_not_dismissable: "To delete an item a person added, use manager_archive.",
+    proposal_replaced: "Read the item with manager_show to see the suggested changes waiting and their ids.",
+    proposal_decided: "Read the item with manager_show for what is still waiting.",
+    timeline_not_live: "Link to a live timeline item of the same client (manager_list with timeline true).",
+    item_not_archived: "It is already in the lists; nothing to bring back.",
+    request_reused: "Add again without request_id to make a new item.",
+};
+/** Fields named for the wrong table: a refusal that says which, so the caller fixes the call. */
+function wrongTable(tool, args, timeline) {
+    const misplaced = (timeline ? ACTION_ONLY : TIMELINE_ONLY).filter((key) => args[key] !== undefined);
+    if (misplaced.length === 0)
+        return null;
+    return errorResult(`Refused (wrong_table_field): ${tool} ${timeline ? "with timeline true" : "on an action"} does not take ${misplaced.join(", ")}`
+        + `${timeline ? " (those are for actions)" : " (those are for timeline items; pass timeline true)"}. Nothing was changed.`);
+}
+/** The item fields a caller named, in the door's own spelling. A field left out stays out: the door keeps its value. */
+function itemFields(args) {
+    const wire = {
+        title: "title", detail: "detail", kind: "kind", due: "dueAt", date: "occurredAt",
+        owners: "owners", owner_name: "ownerName", meeting_id: "timelineItemId", source_url: "sourceUrl", source_note: "sourceNote",
+    };
+    const body = {};
+    for (const [key, name] of Object.entries(wire))
+        if (args[key] !== undefined)
+            body[name] = args[key];
+    return body;
+}
+/** The first `limit` rows of a list, with the cut named (a client with hundreds of items is not shipped whole). */
+function page(rows, limit) {
+    const all = Array.isArray(rows) ? rows : [];
+    return { rows: all.slice(0, limit), more: all.length > limit };
+}
+const LIST_CAP = 500;
+const QUEUE_CAP = 200;
+/** What follows "Showing N of TOTAL" when rows are left out: the server's own cap when it cut, then the next step. */
+function nextStep(serverCap, limit, narrow) {
+    const steps = [...(limit < MAX_LIMIT ? [serverCap ? `raise limit to ${MAX_LIMIT}` : "raise limit"] : []), ...(narrow ? ["narrow with filters"] : [])];
+    return `${serverCap ? ` (server cap ${serverCap})` : ""}${steps.length > 0 ? `; ${steps.join(" or ")}` : ""}.`;
+}
+/** The Manager writes nothing when an action already has the status asked for, so its version comes back as sent. */
+const sameVersion = (answer, version) => answer.item?.updatedAt === version;
+const titleOf = (answer) => {
+    const item = answer.item;
+    return typeof item?.title === "string" ? item.title : "that item";
+};
+export function registerManagerTools(server, deps) {
+    const register = registrarFor(server);
+    /** One door call, relayed: the data on success, the Manager's reason and sentence on a refusal. `suffix` is added when the answer was lost or the server failed (the write may or may not have happened). */
+    async function ask(tool, method, path, body, headline, extra = {}) {
+        return withSession(deps, async (session) => {
+            const response = await callAgentDoor(session, deps.fetchImpl, method, path, body);
+            if (!response.ok)
+                return errorResult(`${managerFailureText(tool, response)}${response.transportError || response.status >= 500 ? (extra.suffix ?? "") : ""}`);
+            const shown = extra.shape ? extra.shape(response.body) : response.body;
+            return dataResult(headline(shown), shown);
+        });
+    }
+    const door = (rest) => `/api/manager/${rest}`;
+    const table = (args) => (args.timeline === true ? "timeline" : "actions");
+    const itemPath = (args, tail = "") => door(`${table(args)}/${encodeURIComponent(String(args.item_id))}${tail}`);
+    const clientPath = (args, tail) => door(`clients/${encodeURIComponent(String(args.client_id))}/${tail}`);
+    const limitOf = (args) => (typeof args.limit === "number" ? args.limit : DEFAULT_LIMIT);
+    const LIMIT = z.number().int().min(1).max(MAX_LIMIT).optional().describe(`How many rows to return, 1 to ${MAX_LIMIT}. Default ${DEFAULT_LIMIT}; the count of every match is always given.`);
+    register("manager_clients", {
+        title: "List Tower Manager clients",
+        description: "The clients Tower Manager follows, each with its id (use it as client_id in the other manager tools), name, onboarding state and counts. Read-only. `cockpit manager clients`, same door.",
+        inputSchema: {},
+    }, async () => ask("manager_clients", "GET", door("clients"), undefined, (answer) => `${Array.isArray(answer.clients) ? answer.clients.length : 0} client(s).`));
+    register("manager_list", {
+        title: "List a client's actions or timeline",
+        description: "One client's actions (to-dos, default) or, with timeline true, its timeline items. Each row carries its id, title, owners, date, review state, any change the AI proposed and `updatedAt` (the `version` a write needs). "
+            + "Filters: needs_review (only items waiting for a person), search, archived (only archived items), review (dismissed or all: dismissed items are hidden by default), plus for actions status (open default, done, dropped, closed, all), owner (one person), no_owner, overdue; for timeline items kind. "
+            + `Returns the first ${DEFAULT_LIMIT} rows by default (limit up to ${MAX_LIMIT}); \`total\` is every match. Read-only. \`cockpit manager list\`, same door.`,
+        inputSchema: {
+            client_id: CLIENT_ID,
+            timeline: TIMELINE,
+            needs_review: z.boolean().optional().describe("true: only items waiting for a person (never reviewed, or carrying a proposed change)."),
+            status: z.enum(ACTION_STATUSES).optional().describe("Actions only: open (default), done, dropped, closed (done and dropped) or all."),
+            owner: z.string().min(1).max(120).optional().describe("Actions only: one owner's name; a shared action is under each of its owners."),
+            no_owner: z.boolean().optional().describe("Actions only: true for actions nobody owns. Not with owner."),
+            overdue: z.boolean().optional().describe("Actions only: true for open actions due before today."),
+            kind: z.enum(KINDS).optional().describe("Timeline items only: meeting, document, milestone or internal."),
+            archived: z.boolean().optional().describe("true: only archived items."),
+            review: z.enum(["dismissed", "all"]).optional().describe("Dismissed items are hidden by default: dismissed lists only those the person dismissed, all lists every item whatever its review state."),
+            search: z.string().min(1).max(200).optional().describe("Words to find in the title or detail."),
+            limit: LIMIT,
+        },
+    }, async (args) => {
+        const timeline = args.timeline === true;
+        const refused = wrongTable("manager_list", args, timeline);
+        if (refused)
+            return refused;
+        const query = new URLSearchParams();
+        if (typeof args.kind === "string")
+            query.set("kind", args.kind);
+        if (typeof args.status === "string")
+            query.set("status", args.status);
+        if (typeof args.owner === "string")
+            query.set("owner", args.owner);
+        if (args.no_owner === true)
+            query.set("noOwner", "true");
+        if (args.overdue === true)
+            query.set("overdue", "true");
+        if (args.needs_review === true)
+            query.set("needsReview", "true");
+        if (args.archived === true)
+            query.set("archived", "true");
+        if (args.review === "dismissed")
+            query.set("review", "rejected"); // Dismiss writes review status "rejected"
+        if (args.review === "all")
+            query.set("review", "all");
+        if (typeof args.search === "string")
+            query.set("q", args.search);
+        const limit = limitOf(args);
+        return ask("manager_list", "GET", clientPath(args, `${table(args)}${queryString(query)}`), undefined, (answer) => answer.shown === answer.total || typeof answer.total !== "number"
+            ? `${String(answer.shown)} of ${String(answer.total ?? answer.shown)} ${timeline ? "timeline item(s)" : "action(s)"}.`
+            : `Showing ${String(answer.shown)} of ${String(answer.total)}${nextStep(answer.truncated === true ? LIST_CAP : null, limit, true)}`, { shape: (answer) => {
+                const { rows, more } = page(answer.items, limit);
+                return { ...answer, items: rows, shown: rows.length, more_exist: more };
+            } });
+    });
+    register("manager_client", {
+        title: "Read one Manager client",
+        description: "One client in full: its name, Notion root, onboarding state, when it was last synced and whether a run holds it now, its latest read summed over its receipts, pages waiting and incomplete, and its newest brief rebuild. "
+            + "Read-only. `cockpit manager client`, same door.",
+        inputSchema: { client_id: CLIENT_ID },
+    }, async (args) => ask("manager_client", "GET", door(`clients/${encodeURIComponent(String(args.client_id))}`), undefined, (answer) => `${typeof answer.name === "string" ? answer.name : "That client"}.`));
+    register("manager_run", {
+        title: "Read one Manager run's receipt",
+        description: "One run's receipt: its kind (sync, extract, onboard, brief_rebuild), when it started and finished, why it ended (reason), and its counts (pages seen, changed, fetched, archived; operations applied, proposed, rejected; model and tokens). "
+            + "Read-only. `cockpit manager run`, same door.",
+        inputSchema: { run_id: z.string().uuid().describe("The run's id: `changedByRunId` on a manager_queue row, `runId` on a proposal in manager_show, or `lastBriefRebuild.runId` in manager_client.") },
+    }, async (args) => ask("manager_run", "GET", door(`runs/${encodeURIComponent(String(args.run_id))}`), undefined, (answer) => `Run ${typeof answer.kind === "string" ? answer.kind : ""} (${typeof answer.reason === "string" ? answer.reason : "no reason named"}), ${answer.finishedAt ? "finished" : "not finished"}.`));
+    register("manager_queue", {
+        title: "A client's review queue",
+        description: "Changes from syncs after onboarding that nobody has reviewed yet, for one client: what a later sync added, changed or proposed, timeline items and actions, each with why it is queued (added, changed, proposal). "
+            + "The onboarding's own items are not in it, so an empty queue does not mean everything is reviewed: for that, use manager_list with needs_review true and status all, and manager_list with timeline true and needs_review true. "
+            + "Work it with manager_approve, manager_dismiss and manager_proposal. Read-only. `cockpit manager queue`, same door.",
+        inputSchema: { client_id: CLIENT_ID, limit: LIMIT },
+    }, async (args) => {
+        const limit = limitOf(args);
+        return ask("manager_queue", "GET", clientPath(args, "review-queue"), undefined, (answer) => {
+            const total = (answer.total ?? {});
+            const shown = (answer.shown ?? {});
+            const [totalTl, totalAc] = [total.timeline ?? 0, total.actions ?? 0];
+            if ((shown.timeline ?? 0) >= totalTl && (shown.actions ?? 0) >= totalAc)
+                return `${totalTl} timeline item(s) and ${totalAc} action(s) in the queue.`;
+            return `Showing ${String(shown.timeline)} of ${totalTl} timeline item(s) and ${String(shown.actions)} of ${totalAc} action(s)${nextStep(answer.truncated === true ? QUEUE_CAP : null, limit, false)}`;
+        }, { shape: (answer) => {
+                const timeline = page(answer.timeline, limit);
+                const actions = page(answer.actions, limit);
+                return { ...answer, timeline: timeline.rows, actions: actions.rows, shown: { timeline: timeline.rows.length, actions: actions.rows.length }, more_exist: timeline.more || actions.more };
+            } });
+    });
+    register("manager_show", {
+        title: "Read one Manager item",
+        description: "One action (default) or, with timeline true, one timeline item in full: its fields, the sources it cites, every change the AI proposed (each with the `id` manager_proposal needs), "
+            + "and what it links to. `updatedAt` is the `version` the next write on it needs. Read-only. `cockpit manager show`, same door.",
+        inputSchema: { item_id: ITEM_ID, timeline: TIMELINE },
+    }, async (args) => ask("manager_show", "GET", itemPath(args), undefined, (answer) => `${titleOf(answer)}.`));
+    register("manager_add", {
+        title: "Add an action or timeline item",
+        description: "Adds one action (default: title, optional detail, due date, owners, and meeting_id to link it to a timeline item) or, with timeline true, one timeline item (kind and title are required; date is optional: "
+            + "without one it lands in Needs a date). Dates are YYYY-MM-DD. The item is the person's own and needs no review; source_url and source_note say where it came from. "
+            + "The answer (and a lost-answer or server-error refusal) carries `request_id`: add again with the same request_id and no duplicate is made; pass your own request_id (a uuid) up front to be able to retry. "
+            + "A request_id already used by another person or client is refused (request_reused). `cockpit manager add`, same door.",
+        inputSchema: {
+            client_id: CLIENT_ID,
+            timeline: TIMELINE,
+            kind: z.enum(KINDS).optional().describe("Timeline items only, required there: meeting, document, milestone or internal."),
+            title: z.string().min(1).max(300).describe("The item's title."),
+            detail: z.string().max(4000).optional().describe("A short paragraph about the item."),
+            due: DATE.optional().describe("Actions only: the due date, YYYY-MM-DD."),
+            date: DATE.optional().describe("Timeline items only: when it happened or will, YYYY-MM-DD. Leave out for undated."),
+            owners: OWNERS,
+            owner_name: OWNER_NAME,
+            meeting_id: z.string().uuid().optional().describe("Actions only: the timeline item (id from manager_list with timeline true) to link the action to."),
+            source_url: z.string().url().max(2000).regex(/^https?:\/\//i).optional().describe("Where it came from: an http(s) link, as the web form's Source. manager_show lists it."),
+            source_note: z.string().min(1).max(1000).optional().describe("Where it came from, in words (at most 1,000 characters). manager_show lists it."),
+            request_id: z.string().uuid().optional().describe("A uuid naming this add, so a retry returns the same item. Default: a fresh one."),
+        },
+    }, async (args) => {
+        const timeline = args.timeline === true;
+        const refused = wrongTable("manager_add", args, timeline);
+        if (refused)
+            return refused;
+        const requestId = typeof args.request_id === "string" ? args.request_id : randomUUID();
+        return ask("manager_add", "POST", clientPath(args, table(args)), { ...itemFields(args), requestId }, (answer) => `Added ${titleOf(answer)}. request_id ${String(answer.request_id)}: retry with it if unsure it was saved.`, { shape: (answer) => ({ ...answer, request_id: requestId }), suffix: ` If unsure whether it was saved, call manager_add again with request_id ${requestId}: no duplicate is made.` });
+    });
+    register("manager_edit", {
+        title: "Edit an action or timeline item",
+        description: "Changes the fields you name on an action (default) or, with timeline true, a timeline item; a field you leave out keeps its value. `version` is required. Dates are YYYY-MM-DD; null clears a date. "
+            + "Actions: owners (leave out to leave the owners alone, [] to clear, a list to replace). Timeline items: kind, date, owner_name. `cockpit manager edit`, same door.",
+        inputSchema: {
+            item_id: ITEM_ID,
+            timeline: TIMELINE,
+            version: VERSION,
+            title: z.string().min(1).max(300).optional().describe("The new title."),
+            detail: z.string().max(4000).optional().describe("The new detail; an empty string clears it."),
+            kind: z.enum(KINDS).optional().describe("Timeline items only: the new kind."),
+            due: DATE.nullable().optional().describe("Actions only: the new due date, YYYY-MM-DD, or null for none."),
+            date: DATE.nullable().optional().describe("Timeline items only: the new date, YYYY-MM-DD, or null for undated."),
+            owners: OWNERS,
+            owner_name: OWNER_NAME,
+        },
+    }, async (args) => {
+        const refused = wrongTable("manager_edit", args, args.timeline === true);
+        if (refused)
+            return refused;
+        return ask("manager_edit", "PATCH", itemPath(args), { ...itemFields(args), expectedUpdatedAt: args.version }, (answer) => `Saved ${titleOf(answer)}.`);
+    });
+    const ACTION_VERSION = { item_id: z.string().uuid().describe("The action's id, from manager_list."), version: VERSION };
+    register("manager_done", {
+        title: "Mark an action done",
+        description: "Ticks an action off (status done). Actions only. `version` is required. manager_reopen undoes it. `cockpit manager done`, same door.",
+        inputSchema: ACTION_VERSION,
+    }, async (args) => ask("manager_done", "POST", itemPath({ ...args, timeline: false }, "/status"), { status: "done", expectedUpdatedAt: args.version }, (answer) => (sameVersion(answer, args.version) ? `Already done; nothing changed: ${titleOf(answer)}.` : `Marked done: ${titleOf(answer)}.`)));
+    register("manager_reopen", {
+        title: "Reopen an action",
+        description: "Puts a done action back to open. Actions only. `version` is required. `cockpit manager reopen`, same door.",
+        inputSchema: ACTION_VERSION,
+    }, async (args) => ask("manager_reopen", "POST", itemPath({ ...args, timeline: false }, "/status"), { status: "open", expectedUpdatedAt: args.version }, (answer) => (sameVersion(answer, args.version) ? `Already open; nothing changed: ${titleOf(answer)}.` : `Reopened: ${titleOf(answer)}.`)));
+    register("manager_link", {
+        title: "Link an action to a timeline item",
+        description: "Links an action to a timeline item (a meeting, say) by meeting_id, or unlinks it when meeting_id is null. Actions only. `version` is the action's. `cockpit manager link`, same door.",
+        inputSchema: {
+            ...ACTION_VERSION,
+            meeting_id: z.string().uuid().nullable().describe("The timeline item's id (manager_list with timeline true), or null to unlink."),
+        },
+    }, async (args) => ask("manager_link", "POST", itemPath({ ...args, timeline: false }, "/link"), { timelineItemId: args.meeting_id, expectedUpdatedAt: args.version }, (answer) => `${args.meeting_id === null ? "Unlinked" : "Linked"}: ${titleOf(answer)}.`));
+    const AS_YOU = "This decides the AI's suggestion as the person whose key this is: their name is recorded as the reviewer, and the item leaves the review queue unless a suggested change still waits on it. It reaches nobody, so no confirm is needed.";
+    register("manager_approve", {
+        title: "Approve an AI-added item",
+        description: `Approves an item the AI added or changed: it stays, marked reviewed. ${AS_YOU} For an item carrying a proposed change use manager_proposal. `
+            + "Approving a DISMISSED item brings it back into the lists (the web's Restore button), so the item is read first: a dismissed one is refused unless restore is true, and restore on an item that is not dismissed is refused too. Nothing is written on a refusal. "
+            + "`version` is required. `cockpit manager approve` (with `--restore`), same door.",
+        inputSchema: {
+            item_id: ITEM_ID,
+            timeline: TIMELINE,
+            version: VERSION,
+            restore: z.boolean().optional().describe("true to bring a DISMISSED item back into the lists, marked reviewed by the person whose key this is. Only for a dismissed item; leave it out to approve an item that is not dismissed."),
+        },
+    }, async (args) => withSession(deps, async (session) => {
+        const read = await callAgentDoor(session, deps.fetchImpl, "GET", itemPath(args), undefined);
+        if (!read.ok)
+            return errorResult(managerFailureText("manager_show", read));
+        const dismissed = read.body.item?.reviewStatus === "rejected";
+        const restore = args.restore === true;
+        if (dismissed && !restore) {
+            return errorResult("Refused (restore_needed): This item is dismissed. Approving it brings it back into the lists, marked reviewed by you. To do that, call manager_approve again with restore true. Nothing was changed.");
+        }
+        if (!dismissed && restore) {
+            return errorResult("Refused (not_dismissed): restore is only for a dismissed item; this one is not dismissed. Call manager_approve without restore. Nothing was changed.");
+        }
+        const response = await callAgentDoor(session, deps.fetchImpl, "POST", itemPath(args, "/review"), { decision: "accept", expectedUpdatedAt: args.version });
+        if (!response.ok)
+            return errorResult(managerFailureText("manager_approve", response));
+        return dataResult(restore ? `Restored: ${titleOf(response.body)}, back in the lists and marked reviewed by you.` : `Approved ${titleOf(response.body)}.`, response.body);
+    }));
+    register("manager_dismiss", {
+        title: "Dismiss an AI-added item",
+        description: `Dismisses an item the AI added that does not belong: it stops showing in the lists. ${AS_YOU} For a proposed change use manager_proposal with decision dismiss. `
+            + "`version` is required. `cockpit manager dismiss`, same door.",
+        inputSchema: { item_id: ITEM_ID, timeline: TIMELINE, version: VERSION },
+    }, async (args) => ask("manager_dismiss", "POST", itemPath(args, "/review"), { decision: "dismiss", expectedUpdatedAt: args.version }, (answer) => `Dismissed ${titleOf(answer)}.`));
+    register("manager_proposal", {
+        title: "Decide one AI-proposed change",
+        description: `Decides one change the AI proposed to an item a person owns: accept applies it, dismiss keeps the item as it is and remembers the refusal, add_new makes the proposal an item of its own (update and duplicate proposals only; a close proposal is refused). ${AS_YOU} `
+            + "`proposal_id` is the `id` of an entry in the item's proposedChanges (manager_show); a newer run may have replaced it, then the decision is refused. `version` is required. `cockpit manager proposal`, same door.",
+        inputSchema: {
+            item_id: ITEM_ID,
+            timeline: TIMELINE,
+            version: VERSION,
+            proposal_id: z.string().min(1).max(64).describe("The proposal's id, from the item's proposedChanges in manager_show."),
+            decision: z.enum(["accept", "dismiss", "add_new"]).describe("accept, dismiss, or add_new (make it a new item)."),
+        },
+    }, async (args) => ask("manager_proposal", "POST", itemPath(args, "/proposal"), {
+        decision: args.decision === "add_new" ? "add_as_new" : args.decision,
+        proposalId: args.proposal_id,
+        expectedUpdatedAt: args.version,
+    }, (answer) => {
+        const added = answer.added;
+        return `Decided (${String(args.decision)}) on ${titleOf(answer)}.${typeof added?.title === "string" ? ` Added as a new item: ${added.title}.` : ""}`;
+    }));
+    register("manager_archive", {
+        title: "Archive an item",
+        description: "Archives an action (default) or timeline item (timeline true): it leaves the lists but is kept, and manager_unarchive puts it back as it was (list archived ones with manager_list archived true). `version` is required. `cockpit manager archive`, same door.",
+        inputSchema: { item_id: ITEM_ID, timeline: TIMELINE, version: VERSION },
+    }, async (args) => ask("manager_archive", "POST", itemPath(args, "/archive"), { expectedUpdatedAt: args.version }, (answer) => `Archived ${titleOf(answer)}.`));
+    register("manager_unarchive", {
+        title: "Unarchive an item",
+        description: "Puts an archived action (default) or timeline item (timeline true) back as it was. `version` is the archived item's updatedAt. `cockpit manager unarchive`, same door.",
+        inputSchema: { item_id: ITEM_ID, timeline: TIMELINE, version: VERSION },
+    }, async (args) => ask("manager_unarchive", "POST", itemPath(args, "/unarchive"), { expectedUpdatedAt: args.version }, (answer) => `Unarchived ${titleOf(answer)}.`));
+}

package/dist/manager-twins.d.ts

@@ -0,0 +1,10 @@
+/**
+ * The Tower Manager's twins in the verb census (BLI-5006).
+ *
+ * `cockpit manager <verb>` and the `manager_*` tools both ride `/api/manager/**`,
+ * which hands the Manager's own answer back unchanged. A verb that works on both
+ * tables (an action, or a timeline item with `--timeline`) names both doors.
+ * Split out of `verb-census.ts`, which sits at the file-size ratchet.
+ */
+import type { McpTwin } from "./verb-census.js";
+export declare const MANAGER_TWINS: Record<string, McpTwin>;

package/dist/manager-twins.js

@@ -0,0 +1,26 @@
+/**
+ * The Tower Manager's twins in the verb census (BLI-5006).
+ *
+ * `cockpit manager <verb>` and the `manager_*` tools both ride `/api/manager/**`,
+ * which hands the Manager's own answer back unchanged. A verb that works on both
+ * tables (an action, or a timeline item with `--timeline`) names both doors.
+ * Split out of `verb-census.ts`, which sits at the file-size ratchet.
+ */
+export const MANAGER_TWINS = {
+    "manager clients": { tool: "manager_clients", door: "GET /api/manager/clients" },
+    "manager client": { tool: "manager_client", door: "GET /api/manager/clients/[clientId]" },
+    "manager run": { tool: "manager_run", door: "GET /api/manager/runs/[runId]" },
+    "manager list": { tool: "manager_list", door: "GET /api/manager/clients/[clientId]/actions (or /timeline)" },
+    "manager queue": { tool: "manager_queue", door: "GET /api/manager/clients/[clientId]/review-queue" },
+    "manager show": { tool: "manager_show", door: "GET /api/manager/actions/[id] (or /timeline/[id])" },
+    "manager add": { tool: "manager_add", door: "POST /api/manager/clients/[clientId]/actions (or /timeline)" },
+    "manager edit": { tool: "manager_edit", door: "PATCH /api/manager/actions/[id] (or /timeline/[id])" },
+    "manager done": { tool: "manager_done", door: "POST /api/manager/actions/[id]/status" },
+    "manager reopen": { tool: "manager_reopen", door: "POST /api/manager/actions/[id]/status" },
+    "manager link": { tool: "manager_link", door: "POST /api/manager/actions/[id]/link" },
+    "manager approve": { tool: "manager_approve", door: "POST /api/manager/actions/[id]/review (or /timeline/[id]/review)" },
+    "manager dismiss": { tool: "manager_dismiss", door: "POST /api/manager/actions/[id]/review (or /timeline/[id]/review)" },
+    "manager proposal": { tool: "manager_proposal", door: "POST /api/manager/actions/[id]/proposal (or /timeline/[id]/proposal)" },
+    "manager archive": { tool: "manager_archive", door: "POST /api/manager/actions/[id]/archive (or /timeline/[id]/archive)" },
+    "manager unarchive": { tool: "manager_unarchive", door: "POST /api/manager/actions/[id]/unarchive (or /timeline/[id]/unarchive)" },
+};

package/dist/notes-account-tools.d.ts

@@ -5,13 +5,18 @@
  * authenticated with this machine's collector device token.
  *
  * ONE DELIBERATE ASYMMETRY WITH THE CLI, recorded in `verb-census.ts`:
- * **`cockpit notes connect` has NO MCP twin.** It carries a notetaker API key,
- * and the CLI takes that on stdin precisely so it never lands in an argument
- * list. An MCP tool argument IS an argument list: it travels through a model's
- * context window and whatever transcript store sits behind it. That is the
- * same ruling `mail add-imap` and `cal add-ical` already carry, for the same
- * credential shape, and connecting a notetaker stays a thing a person does at
- * a terminal or in Settings.
+ * **an API-key `cockpit notes connect` has NO MCP twin.** It carries a
+ * notetaker API key, and the CLI takes that on stdin precisely so it never
+ * lands in an argument list. An MCP tool argument IS an argument list: it
+ * travels through a model's context window and whatever transcript store sits
+ * behind it. That is the same ruling `mail add-imap` and `cal add-ical`
+ * already carry, for the same credential shape.
+ *
+ * Wispr Flow is the exception that proves it (BLI-17973): its sign-in carries
+ * no credential through a tool call. `notes_connect_wisprflow` starts it over
+ * the same `/api/notes/accounts/wisprflow/start` door the Settings button and
+ * `cockpit notes connect --provider wisprflow` call and returns a link for the
+ * PERSON to click; `notes_connect_status` polls the same status door.
  *
  * Same session discipline as `mail-tools.ts`: the session is loaded fresh per
  * call, so a machine with no `cockpit login` pairing still serves this

package/dist/notes-account-tools.js

@@ -5,13 +5,18 @@
  * authenticated with this machine's collector device token.
  *
  * ONE DELIBERATE ASYMMETRY WITH THE CLI, recorded in `verb-census.ts`:
- * **`cockpit notes connect` has NO MCP twin.** It carries a notetaker API key,
- * and the CLI takes that on stdin precisely so it never lands in an argument
- * list. An MCP tool argument IS an argument list: it travels through a model's
- * context window and whatever transcript store sits behind it. That is the
- * same ruling `mail add-imap` and `cal add-ical` already carry, for the same
- * credential shape, and connecting a notetaker stays a thing a person does at
- * a terminal or in Settings.
+ * **an API-key `cockpit notes connect` has NO MCP twin.** It carries a
+ * notetaker API key, and the CLI takes that on stdin precisely so it never
+ * lands in an argument list. An MCP tool argument IS an argument list: it
+ * travels through a model's context window and whatever transcript store sits
+ * behind it. That is the same ruling `mail add-imap` and `cal add-ical`
+ * already carry, for the same credential shape.
+ *
+ * Wispr Flow is the exception that proves it (BLI-17973): its sign-in carries
+ * no credential through a tool call. `notes_connect_wisprflow` starts it over
+ * the same `/api/notes/accounts/wisprflow/start` door the Settings button and
+ * `cockpit notes connect --provider wisprflow` call and returns a link for the
+ * PERSON to click; `notes_connect_status` polls the same status door.
  *
  * Same session discipline as `mail-tools.ts`: the session is loaded fresh per
  * call, so a machine with no `cockpit login` pairing still serves this
@@ -64,9 +69,9 @@ export function registerNotesAccountTools(server, deps) {
         description: "Every notetaker account (Fellow, Circleback or Wispr Flow; Otter and Granola are named but not connectable) "
             + "connected to Tower that you may see: name, provider, workspace, health and when it was last read, plus the feeds the organization reads for "
             + "everyone (Granola, through its own key). Start here when you need an account id. ALSO answers `coverage`: for every member of the team, which notetakers feed "
-            + "their notes, and which members have none (BLI-12926). Connecting one is NOT available here on purpose — it "
+            + "their notes, and which members have none (BLI-12926). Connecting a Fellow or Circleback key is NOT available here on purpose — it "
             + "carries a credential, so it is done at a terminal with `cockpit notes connect --provider fellow --workspace "
-            + "<subdomain> --key-stdin` (or `--provider wisprflow`, which signs in through a browser) or in Tower's Settings.",
+            + "<subdomain> --key-stdin` or in Tower's Settings. Wispr Flow signs in with a link instead: notes_connect_wisprflow.",
         inputSchema: {},
     }, async () => withSession(deps, async (session) => {
         const response = await callAgentDoor(session, deps.fetchImpl, "GET", "/api/notes/accounts");
@@ -155,10 +160,40 @@ export function registerNotesAccountTools(server, deps) {
             accounts,
         });
     }));
+    register("notes_connect_wisprflow", {
+        title: "Start connecting Wispr Flow",
+        description: "Starts signing in to Wispr Flow, the notetaker with no API key. Answers `authorize_url`: give it to the person as a link to click "
+            + "(they approve on Wispr Flow's page, signed in with the Wispr account that holds their notes, using Google, Apple, Microsoft or company single sign-on) "
+            + "and `attempt_id`, then call notes_connect_status with it to see whether it connected. It is the same door the Settings button and "
+            + "`cockpit notes connect --provider wisprflow` call. Carries no credential: the link is one-time and the verifier stays on Tower's server. "
+            + "Refuses with account_already_connected when Wispr Flow is already connected (detach it first).",
+        inputSchema: {},
+    }, async () => withSession(deps, async (session) => {
+        const response = await callAgentDoor(session, deps.fetchImpl, "POST", "/api/notes/accounts/wisprflow/start", {});
+        if (!response.ok)
+            return errorResult(doorFailureText("notes_connect_wisprflow", response));
+        const url = String(response.body.authorize_url ?? "");
+        return textResult(`Open this link in a browser and approve on Wispr Flow's page, then check with notes_connect_status (attempt_id ${String(response.body.attempt_id)}), `
+            + `expires ${String(response.body.expires_at)}:\n${url}`, response.body);
+    }));
+    register("notes_connect_status", {
+        title: "Check a Wispr Flow sign-in",
+        description: "Where a Wispr Flow sign-in stands: pending (still waiting for the click), connected (with the account id), failed (with a reason label) or expired. "
+            + "Pass the attempt_id notes_connect_wisprflow returned. Only the person who started the attempt can read it.",
+        inputSchema: { attempt_id: z.string().uuid().describe("The attempt_id notes_connect_wisprflow returned.") },
+    }, async (args) => withSession(deps, async (session) => {
+        const response = await callAgentDoor(session, deps.fetchImpl, "GET", `/api/notes/accounts/wisprflow/status?attempt_id=${encodeURIComponent(String(args.attempt_id))}`);
+        if (!response.ok)
+            return errorResult(doorFailureText("notes_connect_status", response));
+        const reason = response.body.reason ? ` (${String(response.body.reason)})` : "";
+        return textResult(`Wispr Flow sign-in ${String(response.body.status)}${reason}.`, response.body);
+    }));
     register("notes_sync", {
         title: "Read a notetaker now",
         description: "Reads the meetings this notetaker has recorded since the last pass and files them as meeting notes. "
-            + "Answers with the run's own counts and reason label, which is exactly what the ledger row says. "
+            + "One call is a bounded pass (under a minute): it answers with the run id, the run's own counts and reason label "
+            + "(exactly what the ledger row says), and `more`. While `more` is true, call notes_sync again; each call resumes "
+            + "from the saved cursor, and `done` means the account is caught up. "
             + "The half-hourly cron does this unasked; this is the button for when you do not want to wait.",
         inputSchema: { account_id: z.string().uuid().describe("A notetaker account id (notes_accounts).") },
     }, async (args) => withSession(deps, async (session) => {
@@ -168,6 +203,10 @@ export function registerNotesAccountTools(server, deps) {
         const run = response.body.run;
         return textResult(run
             ? `${run.reason}: ${run.notesAdded} stored, ${run.notesUpdated} unchanged, ${run.notesSkipped} skipped. ${run.detail}`
+                + (run.runId ? ` Run ${run.runId}.` : "")
+                + (run.ok && run.more
+                    ? ` More is waiting${typeof run.remaining === "number" ? ` (${run.remaining})` : ""}: call notes_sync again to continue from the saved cursor.`
+                    : run.ok ? " Caught up." : "")
             : "Tower answered without a run.", response.body);
     }));
     register("notes_detach", {

package/dist/server.js

@@ -43,6 +43,7 @@ import { registerUsageTools } from "./usage-tools.js";
 import { registerLogsTools } from "./logs-tools.js";
 import { registerBrainTools } from "./brain-tools.js";
 import { registerApprovalsTools } from "./approvals-tools.js";
+import { registerManagerTools } from "./manager-tools.js";
 import { registerImageTools } from "./image-tools.js";
 import { registerSourcesTools } from "./sources-tools.js";
 import { TOWER_SERVER_ICONS } from "./server-icons.js";
@@ -526,6 +527,8 @@ export function createServer(deps, options = {}) {
     registerBrainTools(server, { fetchImpl: deps.fetchImpl });
     // BLI-16945: the approvals inbox, over the same /api/approvals doors the page and `cockpit approvals` call.
     registerApprovalsTools(server, { fetchImpl: deps.fetchImpl });
+    // BLI-5006: Tower Manager's clients, actions and timeline, over the same /api/manager door `cockpit manager` calls.
+    registerManagerTools(server, { fetchImpl: deps.fetchImpl });
     // BLI-16977: image descriptions, over the same /api/ambient/images door `cockpit image` calls.
     registerImageTools(server, { fetchImpl: deps.fetchImpl });
     return server;

package/dist/team-write-tools.d.ts

@@ -1,5 +1,5 @@
 /**
- * `team_invite` / `team_role` / `team_device_revoke` MCP tools (BLI-3756 batch
+ * `team_invite` / `team_role` / `team_device_revoke` (and `team_device_scope`, BLI-17978) MCP tools (BLI-3756 batch
  * 2) — the three things an admin does to the team, over the same `/api/team/*`
  * and `/api/ambient/devices/*` doors `cockpit team` calls with the same device
  * token.

package/dist/team-write-tools.js

@@ -1,5 +1,5 @@
 /**
- * `team_invite` / `team_role` / `team_device_revoke` MCP tools (BLI-3756 batch
+ * `team_invite` / `team_role` / `team_device_revoke` (and `team_device_scope`, BLI-17978) MCP tools (BLI-3756 batch
  * 2) — the three things an admin does to the team, over the same `/api/team/*`
  * and `/api/ambient/devices/*` doors `cockpit team` calls with the same device
  * token.
@@ -167,6 +167,28 @@ export function registerTeamWriteTools(server, deps) {
             return errorResult(doorFailureText("team_device_doctor", response));
         return textResult(`Doctor requested for ${String(device.device_name ?? device.device_id)}. Waiting for its next check-in.`, response.body);
     }));
+    register("team_device_scope", {
+        title: "Give the meeting bot's device its scope, or take it away",
+        description: "Grants (`notetaker_bot`) or revokes (`none`) a device's service scope, the same door as `cockpit team device scope`. "
+            + "Without `notetaker_bot` Tower refuses every meeting claim from the bot (bot_scope_required). Super admin only, decided on the server.",
+        inputSchema: {
+            device: z.string().min(1).max(200).describe("A device id or one exact device name."),
+            scope: z.enum(["notetaker_bot", "none"]).describe("`notetaker_bot` to grant, `none` to revoke."),
+        },
+    }, async (args) => withSession(deps, async (session) => {
+        const listed = await listDevices(deps, session);
+        if (!listed.ok)
+            return errorResult(listed.text);
+        const device = findDevice(listed.devices, String(args.device ?? ""));
+        if (!device?.device_id)
+            return errorResult("No unique device matches. Call team_device_list to see ids and names. Nothing was changed.");
+        const response = await callAgentDoor(session, deps.fetchImpl, "POST", `/api/ambient/devices/${encodeURIComponent(device.device_id)}/scope`, { scope: String(args.scope ?? "") }, WRITE_DEADLINE_MS);
+        if (!response.ok)
+            return errorResult(doorFailureText("team_device_scope", response));
+        const name = String(device.device_name ?? device.device_id);
+        return textResult(response.body.changed === false ? `${name} already had that scope. Nothing changed.`
+            : args.scope === "none" ? `Took the service scope away from ${name}.` : `${name} now holds the notetaker_bot scope.`, response.body);
+    }));
     register("team_device_revoke", {
         title: "End a collector device's pairing",
         description: "Revokes one paired machine's token — `cockpit team device revoke`, same door. Irreversible for that token "

package/dist/tool-catalog.d.ts

@@ -10,7 +10,7 @@
  * from `src/` under Node's type stripping, so it may import nothing.
  */
 /** Tools that change nothing. The smoke calls every one. */
-export declare const READ_TOOLS: readonly ["get_ticket_timeline", "get_active_tickets", "docs_list", "docs_tree", "docs_read", "docs_sync_status", "msg_channels", "msg_read", "msg_thread", "work_board", "work_list_issues", "work_get_issue", "work_issue_history", "work_list_projects", "careers_list", "careers_show", "careers_resume_url", "careers_takehome_show", "careers_settings_show", "rename_pending", "github_pr", "setup_read", "speed_read", "speed_seen", "usage_people", "usage_sessions", "mail_accounts", "mail_inbox", "mail_read", "mail_search", "mail_attachment", "mail_sync_status", "cal_today", "cal_week", "cal_next", "cal_find", "cal_calendars", "cal_sync_status", "jarvis_ask", "jarvis_trace", "jarvis_check", "models_list", "models_card", "models_compare", "brief_read", "brief_history", "brief_status", "brief_missed", "notes_folders", "notes_list", "glossary_list", "notes_show", "notes_shelf", "notes_shelves", "notes_place", "notes_accounts", "notes_recordings", "notes_recording", "notes_sync_status", "notes_granola_status", "notes_bot_upcoming", "ops_status", "slack_coverage", "slack_connect", "slack_disconnect", "slack_read", "memory_map", "memory_rules", "memory_knowledge", "memory_profile", "memory_consolidation", "memory_chain", "settings_show", "settings_list", "model_show", "team_members", "team_github_login_add", "team_github_login_remove", "team_github_login_list", "team_device_list", "scout_board", "workbook_read", "tower_search", "sources_check", "logs_digest", "brain_status", "approvals_list", "approvals_show", "approvals_rules", "image_show"];
+export declare const READ_TOOLS: readonly ["get_ticket_timeline", "get_active_tickets", "docs_list", "docs_tree", "docs_read", "docs_sync_status", "msg_channels", "msg_read", "msg_thread", "work_board", "work_list_issues", "work_get_issue", "work_issue_history", "work_list_projects", "careers_list", "careers_show", "careers_resume_url", "careers_takehome_show", "careers_settings_show", "rename_pending", "github_pr", "setup_read", "speed_read", "speed_seen", "usage_people", "usage_sessions", "mail_accounts", "mail_inbox", "mail_read", "mail_search", "mail_attachment", "mail_sync_status", "cal_today", "cal_week", "cal_next", "cal_find", "cal_calendars", "cal_sync_status", "jarvis_ask", "jarvis_trace", "jarvis_check", "models_list", "models_card", "models_compare", "brief_read", "brief_history", "brief_status", "brief_missed", "notes_folders", "notes_list", "glossary_list", "notes_show", "notes_shelf", "notes_shelves", "notes_place", "notes_accounts", "notes_connect_status", "notes_recordings", "notes_recording", "notes_sync_status", "notes_granola_status", "notes_bot_upcoming", "ops_status", "slack_coverage", "slack_connect", "slack_disconnect", "slack_read", "memory_map", "memory_rules", "memory_knowledge", "memory_profile", "memory_consolidation", "memory_chain", "settings_show", "settings_list", "model_show", "team_members", "team_github_login_add", "team_github_login_remove", "team_github_login_list", "team_device_list", "scout_board", "workbook_read", "tower_search", "sources_check", "logs_digest", "brain_status", "approvals_list", "approvals_show", "approvals_rules", "manager_clients", "manager_client", "manager_run", "manager_list", "manager_queue", "manager_show", "image_show"];
 /**
  * Tools that write, but answer a dry run that writes nothing. The smoke calls
  * each one ONLY with its dry-run flag set (`scripts/smoke.ts` says which).
@@ -21,7 +21,7 @@ export declare const DRY_RUN_TOOLS: readonly ["ops_recompile", "memory_requeue"]
  * candidates, share calendars. Never called by the smoke; covered by the unit
  * suites (`write-tools.test.ts` and each family's own).
  */
-export declare const WRITE_TOOLS: readonly ["image_describe", "approvals_create", "approvals_approve", "approvals_edit", "approvals_reject", "approvals_rules_create", "approvals_rules_revoke", "logs_key_mint", "brain_retry", "emit_event", "docs_create", "docs_update", "docs_delete", "docs_link", "msg_create_channel", "msg_dm", "msg_send", "work_create_issue", "work_update_issue", "work_move_issue", "work_comment_issue", "careers_rescreen", "careers_send_takehome", "careers_invite", "careers_grade", "careers_decide", "careers_followup", "careers_interview", "careers_takehome_set", "careers_settings_set", "rename_resolve", "github_merge", "github_retitle", "github_rerun", "mail_send", "mail_draft", "mail_rename", "mail_enable", "mail_disable", "mail_default", "mail_remove", "cal_change", "mail_sync", "mail_detach", "cal_create", "cal_share", "cal_sync", "cal_detach", "cal_delete", "jarvis_dispatch", "glossary_add", "glossary_remove", "notes_mkdir", "notes_rmdir", "notes_rename", "notes_upload", "notes_paste", "notes_share", "notes_unshare", "notes_move", "notes_sync", "notes_detach", "memory_experience", "scout_start", "scout_dismiss", "scout_undo", "brief_rewrite", "brief_correct", "brief_done", "brief_undone", "settings_set", "settings_delete", "model_set", "team_invite", "team_role", "team_grant", "team_revoke", "team_device_revoke", "team_device_doctor"];
+export declare const WRITE_TOOLS: readonly ["image_describe", "approvals_create", "approvals_approve", "approvals_edit", "approvals_reject", "approvals_rules_create", "approvals_rules_revoke", "manager_add", "manager_edit", "manager_done", "manager_reopen", "manager_link", "manager_approve", "manager_dismiss", "manager_proposal", "manager_archive", "manager_unarchive", "logs_key_mint", "brain_retry", "emit_event", "docs_create", "docs_update", "docs_delete", "docs_link", "msg_create_channel", "msg_dm", "msg_send", "work_create_issue", "work_update_issue", "work_move_issue", "work_comment_issue", "careers_rescreen", "careers_send_takehome", "careers_invite", "careers_grade", "careers_decide", "careers_followup", "careers_interview", "careers_takehome_set", "careers_settings_set", "rename_resolve", "github_merge", "github_retitle", "github_rerun", "mail_send", "mail_draft", "mail_rename", "mail_enable", "mail_disable", "mail_default", "mail_remove", "cal_change", "mail_sync", "mail_detach", "cal_create", "cal_share", "cal_sync", "cal_detach", "cal_delete", "jarvis_dispatch", "glossary_add", "glossary_remove", "notes_mkdir", "notes_rmdir", "notes_rename", "notes_upload", "notes_paste", "notes_share", "notes_unshare", "notes_move", "notes_connect_wisprflow", "notes_sync", "notes_detach", "memory_experience", "scout_start", "scout_dismiss", "scout_undo", "brief_rewrite", "brief_correct", "brief_done", "brief_undone", "settings_set", "settings_delete", "model_set", "team_invite", "team_role", "team_grant", "team_revoke", "team_device_revoke", "team_device_doctor", "team_device_scope"];
 /** Reads that spend a model turn. The smoke's `--no-model` leaves these out. */
 export declare const MODEL_SPENDING_TOOLS: readonly ["jarvis_ask", "jarvis_check"];
 /**

package/dist/tool-catalog.js

@@ -67,6 +67,8 @@ export const READ_TOOLS = [
     "notes_shelves",
     "notes_place",
     "notes_accounts",
+    // BLI-17973: where a Wispr Flow sign-in stands (its start is a write, below).
+    "notes_connect_status",
     // BLI-12926: the alternates of a meeting several notetakers recorded.
     "notes_recordings",
     "notes_recording",
@@ -105,6 +107,13 @@ export const READ_TOOLS = [
     "approvals_list",
     "approvals_show",
     "approvals_rules",
+    // BLI-5006: Tower Manager's clients, lists, review queue and one item. Reads only.
+    "manager_clients",
+    "manager_client",
+    "manager_run",
+    "manager_list",
+    "manager_queue",
+    "manager_show",
     // BLI-16977: a stored image description, by sha256. Reads only.
     "image_show",
 ];
@@ -128,6 +137,17 @@ export const WRITE_TOOLS = [
     "approvals_reject",
     "approvals_rules_create",
     "approvals_rules_revoke",
+    // BLI-5006: Tower Manager's writes. Archive is undone by unarchive, so none is destructive.
+    "manager_add",
+    "manager_edit",
+    "manager_done",
+    "manager_reopen",
+    "manager_link",
+    "manager_approve",
+    "manager_dismiss",
+    "manager_proposal",
+    "manager_archive",
+    "manager_unarchive",
     "logs_key_mint",
     "brain_retry",
     "emit_event",
@@ -177,6 +197,8 @@ export const WRITE_TOOLS = [
     "notes_share",
     "notes_unshare",
     "notes_move",
+    // BLI-17973: registers a client at Wispr and records an attempt, so never smoked.
+    "notes_connect_wisprflow",
     "notes_sync",
     "notes_detach",
     "memory_experience",
@@ -196,6 +218,7 @@ export const WRITE_TOOLS = [
     "team_revoke",
     "team_device_revoke",
     "team_device_doctor",
+    "team_device_scope",
 ];
 /** Reads that spend a model turn. The smoke's `--no-model` leaves these out. */
 export const MODEL_SPENDING_TOOLS = ["jarvis_ask", "jarvis_check"];

package/dist/verb-census-spellings.js

@@ -39,6 +39,7 @@ export const TYPED_SPELLING = {
     "team device-list": "team device list",
     "team device-revoke": "team device revoke",
     "team device-doctor": "team device doctor",
+    "team device-scope": "team device scope",
 };
 /** How a census key is written when a person is being told to type it. */
 export function typedSpelling(verb) {

package/dist/verb-census.js

@@ -29,6 +29,8 @@
 import { readFileSync, readdirSync } from "node:fs";
 import { dirname, join, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
+import { DEVICE_TWINS } from "./device-twins.js";
+import { MANAGER_TWINS } from "./manager-twins.js";
 import { IMAGE_TWINS, MEMORY_STORE_TWINS } from "./memory-twins.js";
 const HERE = dirname(fileURLToPath(import.meta.url));
 export const COLLECTOR_COMMANDS_DIR = resolve(HERE, "../../cockpit-local-collector/src/commands");
@@ -349,9 +351,8 @@ export const MCP_TWINS = {
     "notes bot-upcoming": { tool: "notes_bot_upcoming", door: "GET /api/notetaker/upcoming" },
     "notes granola-status": { tool: "notes_granola_status", door: "GET /api/notes/granola-status" },
     "memory log": { tool: "memory_experience", door: "POST /api/memory/experience" },
-    // BLI-4047, BLI-12927: the memory STORE at a terminal. Its twins live on the
-    // second server, `bli-memory` (see `memory-twins.ts`, which also holds the image twins).
-    ...MEMORY_STORE_TWINS, ...IMAGE_TWINS,
+    // BLI-4047, BLI-12927: the memory STORE (twins on the `bli-memory` server) and images: `memory-twins.ts`; BLI-5006: the Manager: `manager-twins.ts`; BLI-17978: the bot scope: `device-twins.ts`.
+    ...MEMORY_STORE_TWINS, ...IMAGE_TWINS, ...MANAGER_TWINS, ...DEVICE_TWINS,
     // BLI-4533: the compiled one-page profile. Its twin is on `bli-tower`, not
     // `bli-memory`: that server's four tools are the vendor-compatible set the
     // Claude Code plugin expects, and this page is ours.
@@ -410,8 +411,7 @@ export const MCP_TWINS = {
     // are.
     search: { tool: "tower_search", door: "GET /api/search" },
     jarvis: { tool: "jarvis_ask", door: "POST /api/jarvis/cli" },
-    // Entered by hand: a mode flag, not an `action` the parser names (see the
-    // header of this file).
+    // Entered by hand: a mode flag, not an `action` the parser names (see the header of this file).
     "jarvis --trace": { tool: "jarvis_trace", door: "GET /api/ops/trace/[id]" },
 };
 // Human command grammar is separate from the mechanical twin census.
@@ -421,7 +421,7 @@ export const TERMINAL_ONLY = {
     "mail add-imap": "carries a Google app password. The CLI reads it from STDIN precisely so it never lands in an argument list; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so attaching a mailbox stays a thing a person does at a terminal (BLI-3708)",
     "cal add-ical": "carries a calendar's SECRET iCal ADDRESS — a permanent, unauthenticated, read-anything-on-that-calendar URL. The CLI reads it from STDIN precisely so it never lands in an argument list; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so attaching a calendar stays a thing a person does at a terminal (BLI-3709)",
     "brief edit": "opens the person's own $EDITOR on this machine and files what they changed; an agent has no editor to open (commands/editor.ts)",
-    "notes connect": "carries a notetaker credential: an API key (Fellow or Circleback), or for Wispr Flow a browser sign-in whose refresh token the CLI hands straight to Tower. The CLI reads a key from STDIN precisely so it never lands in an argument list, and a browser sign-in has no tool-call equivalent at all; an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so connecting a notetaker stays a thing a person does at a terminal or in Tower's Settings (BLI-4383, BLI-12926)",
+    "notes connect": "an API-key connect (Fellow or Circleback) carries a notetaker credential. The CLI reads the key from STDIN precisely so it never lands in an argument list, and an MCP tool argument travels through a model's context window and whatever transcript store sits behind it, so connecting a key stays a thing a person does at a terminal or in Tower's Settings (BLI-4383). Wispr Flow carries no credential through a call: its sign-in is a link the person clicks, started and watched over the same doors by the MCP tools notes_connect_wisprflow and notes_connect_status (BLI-17973)",
     "notes webhook": "carries the VENDOR's webhook signing secret, the key an HMAC over every delivered payload is checked against. Exactly the ruling `notes connect` already carries, for the same credential shape: the CLI reads it from STDIN so it never lands in an argument list, and an MCP tool argument travels through a model's context window and whatever transcript store sits behind it (BLI-4394)",
 };
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bli-cockpit/mcp",
-  "version": "0.1.121",
+  "version": "0.1.125",
   "private": false,
   "description": "bli-tower: an MCP server over BLI Cockpit's agent doors: JARVIS (jarvis_*), documents (docs_*), channels (msg_*), issues (work_*), the daily page (brief_*), meeting notes (notes_*), the ops board (ops_status/slack_*), settings/team/model, Scout and the workbook, plus the legacy event-stream tools (emit_event, get_ticket_timeline, get_active_tickets).",
   "type": "module",