MRTR: drive input_required manually to keep pending-request UX (#1704) (#1719)

* MRTR: drive input_required manually to keep pending-request UX (#1704)

Follow-up to #1627/#1705. Instead of letting the SDK auto-fulfil the modern
(2026-07-28) MRTR retry loop, InspectorClient now drives it itself so the
Inspector keeps its pending-request UX, owns the loop, and shows era-accurate
semantics.

Core (core/mcp/inspectorClient.ts):
- Construct the SDK Client with `inputRequired: { autoFulfill: false }`
  (unconditional; a no-op on legacy, which never returns input_required).
- New `requestWithInputRequired()` driver: issues the request with
  `allowInputRequired` + `withInputRequired(schema)`, and while the result is
  `input_required`, fulfils each embedded request through the SAME pending
  sampling/elicitation plumbing, then retries the original call with
  `inputResponses` + the echoed `requestState` on a fresh id, bounded by
  MRTR_MAX_ROUNDS.
- Route tools/call, prompts/get, and resources/read through the driver.
- Embedded roots/list is auto-answered from configured roots (no pending UX);
  decline/cancel is echoed to the server (not treated as an abort); a tool-call
  abort mid-round settles the pending request and surfaces
  ToolCallCancelledError.
- Extract shared `enqueuePendingSample`/`enqueuePendingElicitation` helpers used
  by both the legacy inbound handlers and the driver; add a per-request `origin`
  ("server-request" | "input-required") to the pending wrappers.

Web UI:
- New `MrtrOriginNote` element shows an "input_required — your answer returns as
  a retry" note on a modern MRTR round; legacy panels are unchanged.
- Thread `origin` App -> PendingClientRequestModal -> the sampling/elicitation
  panels (and the inline panels).

Test servers: new MRTR presets (multi-round, roots, sampling, loop, edge-case
param shaping). Integration tests cover the manual round-trip (asserting the
pending UI pauses and the retry carries inputResponses + requestState on a new
id), multi-round, roots auto-answer, sampling, decline-echoed, maxRounds bound,
abort mid-round, and requestState-only/inputRequests-only param shaping.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XNdjEPKLG637X8YmhDiEk5

* Review: clarify sampling-reject and between-rounds abort comments (#1704)

Address review nits on PR #1719 (comment-only, no behavior change):
- enqueuePendingSample docstring: sampling has no decline/cancel action; the
  panel either sends a result (resolve) or Rejects (reject → fails the call).
  The shared "declined/cancelled resolves" wording was accurate only for the
  elicitation case.
- callToolWithRetries abort comment: note the between-rounds case — when the
  abort lands while an MRTR round awaits an embedded pending request there is no
  in-flight SDK request, so nothing is sent on the wire; the outcome
  (ToolCallCancelledError, no failed-call record) is identical.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XNdjEPKLG637X8YmhDiEk5

* test-servers: add mrtr-showcase config exercising every MRTR preset (#1704)

A single modern-era streamable-HTTP server (port 3111) bundling all six MRTR
presets — mrtr_confirm, mrtr_two_step, mrtr_sample, mrtr_roots, mrtr_edge,
mrtr_loop — for manual testing of the manual-driving pending-request UX across
single-round, multi-round, sampling, roots, param-shaping, and maxRounds paths.
Document it in the README alongside modern-mrtr-http.json.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XNdjEPKLG637X8YmhDiEk5

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -110,7 +110,7 @@ npm run test-servers:build   # (from clients/web) → tsc -p test-servers, emits
 
 The Vite alias `@modelcontextprotocol/inspector-test-server` (in `clients/web/vite.config.ts`) points at `test-servers/build/index.js` so `getTestMcpServerPath()` resolves to a real `.js` path.
 
-A streamable-HTTP server can also serve the **modern (2026-07-28) protocol era** via the SDK's `createMcpHandler` — set `transport.modern` in the JSON config (`true` for dual-era stateless serving, or `{ "legacy": "reject" }` for modern-only strict), or pass `modern` on the `ServerConfig` for an in-process `createTestServerHttp`. This is what lets an Inspector connection negotiating `protocolEra: "auto" | "modern"` reach the modern leg (populated `server/discover`, sessionless). See `test-servers/configs/modern-http.json`. `test-servers/configs/modern-mrtr-http.json` additionally serves the `mrtr_confirm` tool (preset `mrtr_confirm`, `createMrtrTool`) over the modern leg: its handler returns `inputRequired(...)` embedding a form elicitation, so invoking it produces a real MRTR round-trip (`input_required` → the client fulfils the embedded elicitation and retries with a new id → `complete`) — useful for eyeballing the Protocol view's MRTR conversation grouping. (The legacy `collect_elicitation` preset calls `server.elicitInput`, which errors on the 2026-07-28 leg — server→client requests aren't allowed there; MRTR is the modern replacement.)
+A streamable-HTTP server can also serve the **modern (2026-07-28) protocol era** via the SDK's `createMcpHandler` — set `transport.modern` in the JSON config (`true` for dual-era stateless serving, or `{ "legacy": "reject" }` for modern-only strict), or pass `modern` on the `ServerConfig` for an in-process `createTestServerHttp`. This is what lets an Inspector connection negotiating `protocolEra: "auto" | "modern"` reach the modern leg (populated `server/discover`, sessionless). See `test-servers/configs/modern-http.json`. `test-servers/configs/modern-mrtr-http.json` additionally serves the `mrtr_confirm` tool (preset `mrtr_confirm`, `createMrtrTool`) over the modern leg: its handler returns `inputRequired(...)` embedding a form elicitation, so invoking it produces a real MRTR round-trip (`input_required` → the client fulfils the embedded elicitation and retries with a new id → `complete`). The Inspector drives MRTR manually (`inputRequired: { autoFulfill: false }`), so the embedded elicitation pauses at the pending-request modal (tagged "input_required") for you to answer, then the retry completes — useful for eyeballing both that pending-request UX and the Protocol view's MRTR conversation grouping. `test-servers/configs/mrtr-showcase-http.json` bundles every MRTR preset in one modern server for manual testing: `mrtr_confirm` (single round), `mrtr_two_step` (two elicitation rounds via `requestState`), `mrtr_sample` (embedded sampling → the Sampling panel), `mrtr_roots` (embedded `roots/list`, auto-answered silently from configured roots — no modal), `mrtr_edge` (an `inputRequests`-only round then a `requestState`-only round), and `mrtr_loop` (never completes → the `MRTR_MAX_ROUNDS` bound trips). (The legacy `collect_elicitation` preset calls `server.elicitInput`, which errors on the 2026-07-28 leg — server→client requests aren't allowed there; MRTR is the modern replacement.)
 
 ## Building
 

clients/web/src/App.tsx

@@ -3640,6 +3640,7 @@ function App() {
           kind: "sampling",
           id: activeSample.id,
           request: activeSample.request.params,
+          origin: activeSample.origin,
         };
       }
       const activeElicitation = pendingElicitations[0];
@@ -3651,12 +3652,14 @@ function App() {
             id: activeElicitation.id,
             message: params.message,
             url: params.url,
+            origin: activeElicitation.origin,
           };
         }
         return {
           kind: "elicitation-form",
           id: activeElicitation.id,
           request: params,
+          origin: activeElicitation.origin,
         };
       }
       return null;

clients/web/src/components/elements/MrtrOriginNote/MrtrOriginNote.stories.tsx

@@ -0,0 +1,18 @@
+import type { Meta, StoryObj } from "@storybook/react-vite";
+import { MrtrOriginNote } from "./MrtrOriginNote";
+
+const meta: Meta<typeof MrtrOriginNote> = {
+  title: "Elements/MrtrOriginNote",
+  component: MrtrOriginNote,
+};
+
+export default meta;
+type Story = StoryObj<typeof MrtrOriginNote>;
+
+export const InputRequired: Story = {
+  args: { origin: "input-required" },
+};
+
+export const ServerRequest: Story = {
+  args: { origin: "server-request" },
+};

clients/web/src/components/elements/MrtrOriginNote/MrtrOriginNote.test.tsx

@@ -0,0 +1,23 @@
+import { describe, it, expect } from "vitest";
+import { renderWithMantine, screen } from "../../../test/renderWithMantine";
+import { MrtrOriginNote } from "./MrtrOriginNote";
+
+describe("MrtrOriginNote", () => {
+  it("renders the MRTR note for a modern input-required round", () => {
+    renderWithMantine(<MrtrOriginNote origin="input-required" />);
+    expect(screen.getByText("input_required")).toBeInTheDocument();
+    expect(screen.getByText(/sent back as a retry/i)).toBeInTheDocument();
+  });
+
+  it("renders nothing for a legacy server request", () => {
+    renderWithMantine(<MrtrOriginNote origin="server-request" />);
+    expect(screen.queryByText("input_required")).not.toBeInTheDocument();
+    expect(screen.queryByText(/sent back as a retry/i)).not.toBeInTheDocument();
+  });
+
+  it("defaults to rendering nothing when origin is omitted", () => {
+    renderWithMantine(<MrtrOriginNote />);
+    expect(screen.queryByText("input_required")).not.toBeInTheDocument();
+    expect(screen.queryByText(/sent back as a retry/i)).not.toBeInTheDocument();
+  });
+});

clients/web/src/components/elements/MrtrOriginNote/MrtrOriginNote.tsx

@@ -0,0 +1,48 @@
+import { Badge, Group, Text } from "@mantine/core";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
+
+export interface MrtrOriginNoteProps {
+  /**
+   * How the pending request reached the Inspector. `"server-request"` (the
+   * default) is a legacy server→client request and renders nothing, keeping the
+   * panel visually identical to its pre-MRTR form. `"input-required"` is a
+   * modern (2026-07-28) MRTR round and renders the era-accurate note.
+   */
+  origin?: PendingRequestOrigin;
+}
+
+const NoteRow = Group.withProps({
+  gap: "xs",
+  align: "center",
+  wrap: "nowrap",
+});
+
+const NoteText = Text.withProps({
+  size: "xs",
+  c: "dimmed",
+});
+
+/**
+ * Era-accurate note for a pending sampling/elicitation request. On a modern
+ * MRTR round (`origin: "input-required"`, SEP-2322) the request was embedded in
+ * a tool-call/prompt/resource `input_required` result rather than sent as a
+ * server→client JSON-RPC request, and the user's answer is echoed back to the
+ * server as a retry of the original call. Legacy requests
+ * (`origin: "server-request"`) render nothing so their panels are unchanged.
+ */
+export function MrtrOriginNote({
+  origin = "server-request",
+}: MrtrOriginNoteProps) {
+  if (origin !== "input-required") return null;
+  return (
+    <NoteRow>
+      <Badge variant="outline" color="blue">
+        input_required
+      </Badge>
+      <NoteText>
+        The server returned input_required; your answer is sent back as a retry
+        of the original request (MRTR).
+      </NoteText>
+    </NoteRow>
+  );
+}

clients/web/src/components/groups/ElicitationFormPanel/ElicitationFormPanel.tsx

@@ -8,11 +8,13 @@ import {
   Text,
 } from "@mantine/core";
 import type { ElicitRequestFormParams } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
 import {
   hasMissingRequiredFields,
   type InspectorFormSchema,
 } from "../../../utils/jsonUtils";
 import { SchemaForm } from "../SchemaForm/SchemaForm";
+import { MrtrOriginNote } from "../../elements/MrtrOriginNote/MrtrOriginNote";
 
 export interface ElicitationFormPanelProps {
   request: ElicitRequestFormParams;
@@ -24,6 +26,12 @@ export interface ElicitationFormPanelProps {
   onDecline: () => void;
   /** Dismissal without an explicit choice (maps to the spec's `cancel`). */
   onCancel: () => void;
+  /**
+   * How the request reached the Inspector — a legacy server→client request or a
+   * modern MRTR `input_required` round. Drives the era-accurate note; defaults
+   * to `"server-request"` (legacy, note hidden). (#1704)
+   */
+  origin?: PendingRequestOrigin;
   /**
    * A response has been dispatched; lock the actions so a second click can't
    * resolve the request twice (the underlying handler throws if called again).
@@ -66,6 +74,7 @@ export function ElicitationFormPanel({
   onSubmit,
   onDecline,
   onCancel,
+  origin = "server-request",
   busy = false,
 }: ElicitationFormPanelProps) {
   const requestedSchema = request.requestedSchema as InspectorFormSchema;
@@ -74,6 +83,7 @@ export function ElicitationFormPanel({
   return (
     <Stack gap="md">
       <QuotedMessage>{formatQuoted(request.message)}</QuotedMessage>
+      <MrtrOriginNote origin={origin} />
       <Divider />
       <FieldScroll>
         <SchemaForm

clients/web/src/components/groups/ElicitationUrlPanel/ElicitationUrlPanel.tsx

@@ -8,11 +8,19 @@ import {
   Stack,
   Text,
 } from "@mantine/core";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
+import { MrtrOriginNote } from "../../elements/MrtrOriginNote/MrtrOriginNote";
 
 export interface ElicitationUrlPanelProps {
   message: string;
   url: string;
   requestId: string;
+  /**
+   * How the request reached the Inspector — a legacy server→client request or a
+   * modern MRTR `input_required` round. Drives the era-accurate note; defaults
+   * to `"server-request"` (legacy, note hidden). (#1704)
+   */
+  origin?: PendingRequestOrigin;
   /**
    * The URL has been opened and we're waiting for the user to confirm they
    * finished the external flow. Reveals the "I've completed it" action; opening
@@ -65,11 +73,13 @@ export function ElicitationUrlPanel({
   onOpenInBrowser,
   onComplete,
   onCancel,
+  origin = "server-request",
   busy = false,
 }: ElicitationUrlPanelProps) {
   return (
     <Stack gap="md">
       <ItalicMessage>{message}</ItalicMessage>
+      <MrtrOriginNote origin={origin} />
       <Divider />
       <Text size="sm">The server is requesting you visit:</Text>
       <WrappingCode>{url}</WrappingCode>

clients/web/src/components/groups/InlineElicitationRequest/InlineElicitationRequest.test.tsx

@@ -41,6 +41,21 @@ describe("InlineElicitationRequest", () => {
     expect(screen.getByText("elicitation/create (form)")).toBeInTheDocument();
     expect(screen.getByText(formRequest.message)).toBeInTheDocument();
     expect(screen.getByText("1 of 1")).toBeInTheDocument();
+    // Default origin is a legacy server request — no MRTR note.
+    expect(screen.queryByText("input_required")).not.toBeInTheDocument();
+  });
+
+  it("shows the MRTR note for a modern input_required round", () => {
+    renderWithMantine(
+      <InlineElicitationRequest
+        {...baseProps}
+        request={formRequest}
+        values={{}}
+        origin="input-required"
+      />,
+    );
+    expect(screen.getByText("input_required")).toBeInTheDocument();
+    expect(screen.getByText(/sent back as a retry/i)).toBeInTheDocument();
   });
 
   it("renders schema fields and a Submit button in form mode", () => {

clients/web/src/components/groups/InlineElicitationRequest/InlineElicitationRequest.tsx

@@ -12,14 +12,22 @@ import type {
   ElicitRequest,
   ElicitRequestFormParams,
 } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
 import type { InspectorFormSchema } from "../../../utils/jsonUtils";
 import { SchemaForm } from "../SchemaForm/SchemaForm";
+import { MrtrOriginNote } from "../../elements/MrtrOriginNote/MrtrOriginNote";
 
 export interface InlineElicitationRequestProps {
   request: ElicitRequest["params"];
   queuePosition: string;
   values?: Record<string, unknown>;
   isWaiting?: boolean;
+  /**
+   * How the request reached the Inspector — a legacy server→client request or a
+   * modern MRTR `input_required` round. Drives the era-accurate note; defaults
+   * to `"server-request"` (legacy, note hidden). (#1704)
+   */
+  origin?: PendingRequestOrigin;
   onChange: (values: Record<string, unknown>) => void;
   onSubmit: () => void;
   onCancel: () => void;
@@ -73,6 +81,7 @@ export function InlineElicitationRequest({
   queuePosition,
   values,
   isWaiting,
+  origin = "server-request",
   onChange,
   onSubmit,
   onCancel,
@@ -86,6 +95,7 @@ export function InlineElicitationRequest({
         </Group>
 
         <ItalicMessage>{request.message}</ItalicMessage>
+        <MrtrOriginNote origin={origin} />
 
         {isFormMode(request) && (
           <SchemaForm

clients/web/src/components/groups/InlineSamplingRequest/InlineSamplingRequest.test.tsx

@@ -84,6 +84,20 @@ describe("InlineSamplingRequest", () => {
     expect(screen.getByText("sampling/createMessage")).toBeInTheDocument();
     expect(screen.getByText("1 of 1")).toBeInTheDocument();
     expect(screen.getByText("Please analyze this code.")).toBeInTheDocument();
+    // Default origin is a legacy server request — no MRTR note.
+    expect(screen.queryByText("input_required")).not.toBeInTheDocument();
+  });
+
+  it("shows the MRTR note for a modern input_required round", () => {
+    renderWithMantine(
+      <InlineSamplingRequest
+        {...baseProps}
+        request={textRequest}
+        origin="input-required"
+      />,
+    );
+    expect(screen.getByText("input_required")).toBeInTheDocument();
+    expect(screen.getByText(/sent back as a retry/i)).toBeInTheDocument();
   });
 
   it("renders model hints when provided", () => {

clients/web/src/components/groups/InlineSamplingRequest/InlineSamplingRequest.tsx

@@ -3,11 +3,19 @@ import type {
   CreateMessageRequestParams,
   CreateMessageResult,
 } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
+import { MrtrOriginNote } from "../../elements/MrtrOriginNote/MrtrOriginNote";
 
 export interface InlineSamplingRequestProps {
   request: CreateMessageRequestParams;
   queuePosition: string;
   draftResult?: CreateMessageResult;
+  /**
+   * How the request reached the Inspector — a legacy server→client request or a
+   * modern MRTR `input_required` round. Drives the era-accurate note; defaults
+   * to `"server-request"` (legacy, note hidden). (#1704)
+   */
+  origin?: PendingRequestOrigin;
   onAutoRespond: () => void;
   onEditAndSend: () => void;
   onReject: () => void;
@@ -94,6 +102,7 @@ export function InlineSamplingRequest({
   request,
   queuePosition,
   draftResult,
+  origin = "server-request",
   onAutoRespond,
   onEditAndSend,
   onReject,
@@ -111,6 +120,8 @@ export function InlineSamplingRequest({
           <QueueLabel>{queuePosition}</QueueLabel>
         </Group>
 
+        <MrtrOriginNote origin={origin} />
+
         {modelHints && <Text size="sm">{formatModelHints(modelHints)}</Text>}
 
         <PreviewText>{messagePreview}</PreviewText>

clients/web/src/components/groups/PendingClientRequestModal/PendingClientRequestModal.stories.tsx

@@ -112,7 +112,12 @@ type Story = StoryObj<typeof PendingClientRequestModal>;
 
 export const Sampling: Story = {
   args: {
-    request: { kind: "sampling", id: "sampling-1", request: samplingRequest },
+    request: {
+      kind: "sampling",
+      id: "sampling-1",
+      request: samplingRequest,
+      origin: "server-request",
+    },
   },
   play: async ({ canvasElement, args }) => {
     const body = within(canvasElement.ownerDocument.body);
@@ -125,12 +130,32 @@ export const Sampling: Story = {
   },
 };
 
+// A modern MRTR round: the sampling request was embedded in a tool-call
+// `input_required` result, so the panel shows the era-accurate note.
+export const SamplingInputRequired: Story = {
+  args: {
+    request: {
+      kind: "sampling",
+      id: "sampling-mrtr",
+      request: samplingRequest,
+      origin: "input-required",
+    },
+  },
+  play: async ({ canvasElement }) => {
+    const body = within(canvasElement.ownerDocument.body);
+    await body.findByText("Sampling Request");
+    expect(body.getByText("input_required")).toBeInTheDocument();
+    expect(body.getByText(/sent back as a retry/i)).toBeInTheDocument();
+  },
+};
+
 export const SamplingTall: Story = {
   args: {
     request: {
       kind: "sampling",
       id: "sampling-tall",
       request: tallSamplingRequest,
+      origin: "server-request",
     },
   },
 };
@@ -141,6 +166,7 @@ export const ElicitationForm: Story = {
       kind: "elicitation-form",
       id: "elicitation-1",
       request: formRequest,
+      origin: "server-request",
     },
   },
   play: async ({ canvasElement, args }) => {
@@ -163,6 +189,7 @@ export const ElicitationFormTall: Story = {
       kind: "elicitation-form",
       id: "elicitation-tall",
       request: tallFormRequest,
+      origin: "server-request",
     },
   },
 };
@@ -174,6 +201,7 @@ export const ElicitationUrl: Story = {
       id: "elicitation-2",
       message: "Authorize access in your browser to continue.",
       url: "https://example.com/authorize?token=abc123",
+      origin: "server-request",
     },
   },
   play: async ({ canvasElement }) => {

clients/web/src/components/groups/PendingClientRequestModal/PendingClientRequestModal.test.tsx

@@ -32,12 +32,14 @@ const samplingContent: PendingClientRequestContent = {
   kind: "sampling",
   id: "sampling-1",
   request: samplingRequest,
+  origin: "server-request",
 };
 
 const formContent: PendingClientRequestContent = {
   kind: "elicitation-form",
   id: "elicitation-1",
   request: formRequest,
+  origin: "server-request",
 };
 
 const urlContent: PendingClientRequestContent = {
@@ -45,6 +47,7 @@ const urlContent: PendingClientRequestContent = {
   id: "elicitation-2",
   message: "Authorize access in your browser.",
   url: "https://example.com/authorize",
+  origin: "server-request",
 };
 
 const baseProps = {
@@ -77,6 +80,19 @@ describe("PendingClientRequestModal", () => {
     expect(
       screen.getByText("The server is requesting an LLM completion."),
     ).toBeInTheDocument();
+    // Legacy server-request origin: no MRTR note.
+    expect(screen.queryByText("input_required")).not.toBeInTheDocument();
+  });
+
+  it("shows the MRTR note when the request is a modern input_required round", () => {
+    renderWithMantine(
+      <PendingClientRequestModal
+        {...baseProps}
+        request={{ ...samplingContent, origin: "input-required" }}
+      />,
+    );
+    expect(screen.getByText("input_required")).toBeInTheDocument();
+    expect(screen.getByText(/sent back as a retry/i)).toBeInTheDocument();
   });
 
   it("sends the default stub sampling result when the draft is untouched", async () => {
@@ -176,6 +192,7 @@ describe("PendingClientRequestModal", () => {
           },
         },
       },
+      origin: "server-request",
     };
     renderWithMantine(
       <PendingClientRequestModal {...baseProps} request={formWithDefaults} />,

clients/web/src/components/groups/PendingClientRequestModal/PendingClientRequestModal.tsx

@@ -6,6 +6,7 @@ import type {
   ElicitRequestFormParams,
   ElicitResult,
 } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
 import { SamplingRequestPanel } from "../SamplingRequestPanel/SamplingRequestPanel";
 import { ElicitationFormPanel } from "../ElicitationFormPanel/ElicitationFormPanel";
 import { ElicitationUrlPanel } from "../ElicitationUrlPanel/ElicitationUrlPanel";
@@ -17,12 +18,30 @@ import {
 /**
  * The server-initiated request currently shown in the modal. `id` is the
  * client-side request id; it keys the modal body so per-request draft state
- * resets when the active request changes.
+ * resets when the active request changes. `origin` distinguishes a legacy
+ * server→client request from a modern MRTR `input_required` round so the panels
+ * can show era-accurate semantics (#1704).
  */
 export type PendingClientRequestContent =
-  | { kind: "sampling"; id: string; request: CreateMessageRequestParams }
+  | {
-  | { kind: "elicitation-form"; id: string; request: ElicitRequestFormParams }
+      kind: "sampling";
-  | { kind: "elicitation-url"; id: string; message: string; url: string };
+      id: string;
+      request: CreateMessageRequestParams;
+      origin: PendingRequestOrigin;
+    }
+  | {
+      kind: "elicitation-form";
+      id: string;
+      request: ElicitRequestFormParams;
+      origin: PendingRequestOrigin;
+    }
+  | {
+      kind: "elicitation-url";
+      id: string;
+      message: string;
+      url: string;
+      origin: PendingRequestOrigin;
+    };
 
 export interface PendingClientRequestModalProps {
   /** The active request to display, or null when nothing is pending. */
@@ -103,10 +122,12 @@ function useRespondOnce(): {
 
 function SamplingModalBody({
   request,
+  origin,
   onRespond,
   onReject,
 }: {
   request: CreateMessageRequestParams;
+  origin: PendingRequestOrigin;
   onRespond: (result: CreateMessageResult) => void;
   onReject: () => void;
 }) {
@@ -117,6 +138,7 @@ function SamplingModalBody({
   return (
     <SamplingRequestPanel
       request={request}
+      origin={origin}
       draftResult={draftResult}
       onResultChange={setDraftResult}
       onSend={once(() => onRespond(draftResult))}
@@ -129,10 +151,12 @@ function SamplingModalBody({
 function ElicitationFormModalBody({
   request,
   serverName,
+  origin,
   onRespond,
 }: {
   request: ElicitRequestFormParams;
   serverName: string;
+  origin: PendingRequestOrigin;
   onRespond: (result: ElicitResult) => void;
 }) {
   // Seed with the schema's defaults so default-only fields the user never
@@ -147,6 +171,7 @@ function ElicitationFormModalBody({
     <ElicitationFormPanel
       request={request}
       serverName={serverName}
+      origin={origin}
       values={values}
       onChange={setValues}
       onSubmit={once(() =>
@@ -166,11 +191,13 @@ function ElicitationUrlModalBody({
   id,
   message,
   url,
+  origin,
   onRespond,
 }: {
   id: string;
   message: string;
   url: string;
+  origin: PendingRequestOrigin;
   onRespond: (result: ElicitResult) => void;
 }) {
   const [isWaiting, setIsWaiting] = useState(false);
@@ -179,6 +206,7 @@ function ElicitationUrlModalBody({
     <ElicitationUrlPanel
       message={message}
       url={url}
+      origin={origin}
       requestId={id}
       isWaiting={isWaiting}
       onCopyUrl={() => {
@@ -237,6 +265,7 @@ export function PendingClientRequestModal({
         <SamplingModalBody
           key={request.id}
           request={request.request}
+          origin={request.origin}
           onRespond={onSamplingRespond}
           onReject={onSamplingReject}
         />
@@ -246,6 +275,7 @@ export function PendingClientRequestModal({
           key={request.id}
           request={request.request}
           serverName={serverName}
+          origin={request.origin}
           onRespond={onElicitationRespond}
         />
       )}
@@ -255,6 +285,7 @@ export function PendingClientRequestModal({
           id={request.id}
           message={request.message}
           url={request.url}
+          origin={request.origin}
           onRespond={onElicitationRespond}
         />
       )}

clients/web/src/components/groups/SamplingRequestPanel/SamplingRequestPanel.tsx

@@ -17,7 +17,9 @@ import type {
   CreateMessageRequestParams,
   CreateMessageResult,
 } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "@inspector/core/mcp/types.js";
 import { MessageBubble } from "../../elements/MessageBubble/MessageBubble";
+import { MrtrOriginNote } from "../../elements/MrtrOriginNote/MrtrOriginNote";
 
 export interface SamplingRequestPanelProps {
   request: CreateMessageRequestParams;
@@ -25,6 +27,12 @@ export interface SamplingRequestPanelProps {
   onResultChange: (result: CreateMessageResult) => void;
   onSend: () => void;
   onReject: () => void;
+  /**
+   * How the request reached the Inspector — a legacy server→client request or a
+   * modern MRTR `input_required` round. Drives the era-accurate note; defaults
+   * to `"server-request"` (legacy, note hidden). (#1704)
+   */
+  origin?: PendingRequestOrigin;
   /**
    * A response has been dispatched; lock the actions so a second click can't
    * resolve the request twice (the underlying handler throws if called again).
@@ -90,6 +98,7 @@ export function SamplingRequestPanel({
   onResultChange,
   onSend,
   onReject,
+  origin = "server-request",
   busy = false,
 }: SamplingRequestPanelProps) {
   const {
@@ -110,6 +119,7 @@ export function SamplingRequestPanel({
       <ContentScroll>
         <ContentStack>
           <HintText>The server is requesting an LLM completion.</HintText>
+          <MrtrOriginNote origin={origin} />
 
           <SectionTitle>Messages:</SectionTitle>
           {messages.map((message, index) => (

clients/web/src/test/core/mcp/inspectorClientUrlElicitation.test.ts

@@ -42,6 +42,10 @@ type FakeClient = {
  * we can drive `callTool` through the URL-elicitation error path without a live
  * server. The client is never connected; `callTool` only needs the injected
  * `callTool`/`request` methods plus the (connection-independent) helpers.
+ *
+ * The tool-call path issues `client.request({ method: "tools/call", … })`
+ * (through the MRTR driver, #1704), so these fakes drive the error path from
+ * `request`; `callTool` is left as an unused stub.
  */
 function makeClient(fake: FakeClient): InspectorClient {
   const client = new InspectorClient(
@@ -72,14 +76,14 @@ describe("InspectorClient URL-elicitation error path", () => {
   it("surfaces the URL elicitation, then retries the call on accept", async () => {
     let attempt = 0;
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         attempt += 1;
         if (attempt === 1) {
           throw new UrlElicitationRequiredError([elicitation]);
         }
         return okResult;
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
 
@@ -99,7 +103,7 @@ describe("InspectorClient URL-elicitation error path", () => {
 
     const invocation = await pending;
     expect(invocation.success).toBe(true);
-    expect(fake.callTool).toHaveBeenCalledTimes(2);
+    expect(fake.request).toHaveBeenCalledTimes(2);
     expect(client.getPendingElicitations()).toHaveLength(0);
   });
 
@@ -111,14 +115,14 @@ describe("InspectorClient URL-elicitation error path", () => {
     };
     let attempt = 0;
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         attempt += 1;
         if (attempt === 1) {
           throw new UrlElicitationRequiredError([elicitation, second]);
         }
         return okResult;
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
 
@@ -143,15 +147,15 @@ describe("InspectorClient URL-elicitation error path", () => {
 
     const invocation = await pending;
     expect(invocation.success).toBe(true);
-    expect(fake.callTool).toHaveBeenCalledTimes(2);
+    expect(fake.request).toHaveBeenCalledTimes(2);
   });
 
   it("aborts the call (no retry) when the user cancels a required elicitation", async () => {
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         throw new UrlElicitationRequiredError([elicitation]);
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
     const failedDispatches = trackFailedDispatches(client);
@@ -163,7 +167,7 @@ describe("InspectorClient URL-elicitation error path", () => {
     await client.getPendingElicitations()[0].respond({ action: "cancel" });
 
     await expect(pending).rejects.toThrow(/cancelled/i);
-    expect(fake.callTool).toHaveBeenCalledTimes(1);
+    expect(fake.request).toHaveBeenCalledTimes(1);
     // The abort records exactly one failed history entry — not zero, not a
     // duplicate from the generic catch.
     expect(failedDispatches()).toBe(1);
@@ -172,10 +176,10 @@ describe("InspectorClient URL-elicitation error path", () => {
   it("aborts with a loop error when the server re-requests a completed URL", async () => {
     // The server keeps returning the same URL after the user completes it.
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         throw new UrlElicitationRequiredError([elicitation]);
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
     const failedDispatches = trackFailedDispatches(client);
@@ -192,7 +196,7 @@ describe("InspectorClient URL-elicitation error path", () => {
     await expect(pending).rejects.toBeInstanceOf(UrlElicitationLoopError);
     await expect(pending).rejects.toMatchObject({ url: elicitation.url });
     // Only the initial call + one retry ran; the URL was presented just once.
-    expect(fake.callTool).toHaveBeenCalledTimes(2);
+    expect(fake.request).toHaveBeenCalledTimes(2);
     expect(client.getPendingElicitations()).toHaveLength(0);
     // The loop abort records exactly one failed history entry (the accepted
     // first round records nothing).
@@ -201,10 +205,10 @@ describe("InspectorClient URL-elicitation error path", () => {
 
   it("settles a pending error-path elicitation as cancelled on disconnect (no hang)", async () => {
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         throw new UrlElicitationRequiredError([elicitation]);
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
 
@@ -223,13 +227,13 @@ describe("InspectorClient URL-elicitation error path", () => {
 
   it("rethrows a -32042 error with no elicitations without queuing anything", async () => {
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         throw new ProtocolError(
           ProtocolErrorCode.UrlElicitationRequired,
           "This request requires browser-based authorization.",
         );
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
 
@@ -237,15 +241,15 @@ describe("InspectorClient URL-elicitation error path", () => {
       code: ProtocolErrorCode.UrlElicitationRequired,
     });
     expect(client.getPendingElicitations()).toHaveLength(0);
-    expect(fake.callTool).toHaveBeenCalledTimes(1);
+    expect(fake.request).toHaveBeenCalledTimes(1);
   });
 
   it("rethrows an ordinary error unchanged", async () => {
     const fake: FakeClient = {
-      callTool: vi.fn(async () => {
+      callTool: vi.fn(),
+      request: vi.fn(async () => {
         throw new Error("boom");
       }),
-      request: vi.fn(),
     };
     const client = makeClient(fake);
     const failed = vi.fn();

clients/web/src/test/integration/mcp/inspectorClient-coverage-backfill.test.ts

@@ -567,12 +567,14 @@ describe("InspectorClient coverage backfill", () => {
     it("callTool error path covers general-only, tool-specific-only, and both metadata", async () => {
       client = stdioClient();
       await client.connect();
-      // Force the underlying SDK callTool to throw so the dispatchFailedToolCall
+      // Force the underlying SDK request to throw so the dispatchFailedToolCall
-      // error path runs for each metadata combination.
+      // error path runs for each metadata combination. The tool-call path issues
+      // `client.request({ method: "tools/call", … })` through the MRTR driver
+      // (#1704), so the throw is injected on `request`.
       const c = client as unknown as {
-        client: { callTool: () => Promise<never> };
+        client: { request: () => Promise<never> };
       };
-      c.client.callTool = async () => {
+      c.client.request = async () => {
         throw new Error("forced-call-failure");
       };
 

clients/web/src/test/integration/mcp/inspectorClient-modern-era.test.ts

@@ -1,6 +1,7 @@
 import { describe, it, expect, afterEach } from "vitest";
 import { InspectorClient } from "@inspector/core/mcp/inspectorClient.js";
 import { createTransportNode } from "@inspector/core/mcp/node/transport.js";
+import { ToolCallCancelledError } from "@inspector/core/mcp/toolCallCancelledError.js";
 import {
   eraToVersionNegotiation,
   MODERN_PROTOCOL_VERSION,
@@ -11,9 +12,17 @@ import {
   createTestServerInfo,
   createEchoTool,
   createMrtrTool,
+  createMrtrMultiRoundTool,
+  createMrtrRootsTool,
+  createMrtrSamplingTool,
+  createMrtrLoopTool,
+  createMrtrEdgeCaseTool,
 } from "@modelcontextprotocol/inspector-test-server";
 import type { ServerConfig } from "@modelcontextprotocol/inspector-test-server";
-import type { ContentBlock } from "@modelcontextprotocol/client";
+import type {
+  ContentBlock,
+  JSONRPCRequest,
+} from "@modelcontextprotocol/client";
 
 /**
  * Live coverage of the modern (2026-07-28) connection path (#1700). The bundled
@@ -126,25 +135,55 @@ describe("modern-era negotiation (2026-07-28)", () => {
     expect("text" in content[0] && content[0].text).toContain("hi");
   });
 
-  it("completes an MRTR round-trip (input_required elicitation → retry → final result)", async () => {
+  // Collect every outbound `tools/call` request frame the transport captured,
-    // The modern (2026-07-28) leg has no server→client requests; a tool that
+  // so a test can assert what the MRTR retry actually put on the wire.
-    // needs input returns `input_required` embedding the request, and the client
+  function collectToolCallRequests(
-    // fulfils it and retries with a new id. `createMrtrTool` returns
+    connected: InspectorClient,
-    // `inputRequired({ inputRequests: { confirm: elicit(...) } })` on the first
+  ): JSONRPCRequest[] {
-    // call and the final result once the answer is echoed back.
+    const frames: JSONRPCRequest[] = [];
+    connected.addEventListener("message", (event) => {
+      const entry = event.detail;
+      if (
+        entry.direction === "request" &&
+        "method" in entry.message &&
+        entry.message.method === "tools/call"
+      ) {
+        frames.push(entry.message as JSONRPCRequest);
+      }
+    });
+    return frames;
+  }
+
+  async function startMrtrServer(
+    tool: ReturnType<typeof createMrtrTool>,
+  ): Promise<TestServerHttp> {
     const started = createTestServerHttp({
       serverInfo: createTestServerInfo("modern-mrtr-test", "1.0.0"),
-      tools: [createMrtrTool()],
+      tools: [tool],
       modern: {},
     });
     await started.start();
     server = started;
+    return started;
+  }
 
+  it("drives an MRTR round-trip manually (pauses at the pending UI, retries with inputResponses + requestState on a new id)", async () => {
+    // The modern (2026-07-28) leg has no server→client requests; a tool that
+    // needs input returns `input_required` embedding the request. With the SDK's
+    // auto-fulfil disabled, InspectorClient drives the loop itself: it surfaces
+    // the embedded elicitation through the pending-request UI, then retries the
+    // original call with the answer.
+    const started = await startMrtrServer(createMrtrTool());
     const connected = await connectWithEra(started.url, "modern");
+    const toolCallFrames = collectToolCallRequests(connected);
 
-    // Auto-fulfil the embedded elicitation the MRTR tool requests, the same way
+    // Prove the pending UX actually paused: record that an elicitation was
-    // the UI's pending-request panel would — which drives the SDK's retry.
+    // surfaced BEFORE we answer it (the manual driver enqueued it), then answer.
+    let pausedAtPendingUi = false;
     connected.addEventListener("newPendingElicitation", (event) => {
+      pausedAtPendingUi = true;
+      // The manual driver tags an MRTR round's request "input-required".
+      expect(event.detail.origin).toBe("input-required");
       void event.detail.respond({
         action: "accept",
         content: { confirm: true },
@@ -157,12 +196,213 @@ describe("modern-era negotiation (2026-07-28)", () => {
 
     const result = await connected.callTool(mrtr!, { action: "deploy" });
     expect(result.success).toBe(true);
+    expect(pausedAtPendingUi).toBe(true);
+
     const content = result.result!.content as ContentBlock[];
     const text = content[0] && "text" in content[0] ? content[0].text : "";
     // The final result is only reachable after the input_required round was
     // fulfilled and the original call retried — i.e. the full MRTR loop ran.
     expect(text).toContain("MRTR complete");
     expect(text).toContain("confirm");
+
+    // Two tools/call frames: the original and the retry. The retry has a
+    // DIFFERENT json-rpc id and carries inputResponses + the echoed requestState.
+    expect(toolCallFrames.length).toBe(2);
+    const [original, retry] = toolCallFrames;
+    expect(retry.id).not.toBe(original.id);
+    const retryParams = retry.params as {
+      inputResponses?: Record<string, unknown>;
+      requestState?: unknown;
+    };
+    expect(retryParams.inputResponses?.confirm).toEqual({
+      action: "accept",
+      content: { confirm: true },
+    });
+    // The original call carries no requestState; the retry echoes the opaque
+    // server-minted token verbatim (shape: `mrtr:<action>:<n>`).
+    expect(
+      (original.params as { requestState?: unknown }).requestState,
+    ).toBeUndefined();
+    expect(retryParams.requestState).toMatch(/^mrtr:deploy:/);
+  });
+
+  it("drives a multi-round MRTR (two embedded elicitations in sequence, then completes)", async () => {
+    const started = await startMrtrServer(createMrtrMultiRoundTool());
+    const connected = await connectWithEra(started.url, "modern");
+
+    const seenMessages: string[] = [];
+    connected.addEventListener("newPendingElicitation", (event) => {
+      seenMessages.push(event.detail.request.params.message ?? "");
+      void event.detail.respond({
+        action: "accept",
+        content: { value: `answer-${seenMessages.length}` },
+      });
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_two_step");
+    const result = await connected.callTool(tool!, {});
+    expect(result.success).toBe(true);
+
+    // Both rounds surfaced, in order.
+    expect(seenMessages).toEqual([
+      "Step 1: enter the first value",
+      "Step 2: enter the second value",
+    ]);
+    const content = result.result!.content as ContentBlock[];
+    const text = content[0] && "text" in content[0] ? content[0].text : "";
+    expect(text).toContain("MRTR two-step complete");
+  });
+
+  it("auto-answers an embedded roots/list request from configured roots (no pending UI)", async () => {
+    const started = await startMrtrServer(createMrtrRootsTool());
+    const connected = new InspectorClient(
+      { type: "streamable-http", url: started.url },
+      {
+        environment: { transport: createTransportNode },
+        versionNegotiation: eraToVersionNegotiation("modern"),
+        roots: [{ uri: "file:///workspace", name: "workspace" }],
+      },
+    );
+    client = connected;
+    await connected.connect();
+
+    let surfacedPending = false;
+    connected.addEventListener("newPendingElicitation", () => {
+      surfacedPending = true;
+    });
+    connected.addEventListener("newPendingSample", () => {
+      surfacedPending = true;
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_roots");
+    const result = await connected.callTool(tool!, {});
+    expect(result.success).toBe(true);
+    // roots is answered silently — nothing pauses at the pending UI.
+    expect(surfacedPending).toBe(false);
+    const content = result.result!.content as ContentBlock[];
+    const text = content[0] && "text" in content[0] ? content[0].text : "";
+    expect(text).toContain("client reported 1 root");
+  });
+
+  it("surfaces an embedded sampling request through the pending-sample UI", async () => {
+    const started = await startMrtrServer(createMrtrSamplingTool());
+    const connected = await connectWithEra(started.url, "modern");
+
+    let sampleOrigin: string | undefined;
+    connected.addEventListener("newPendingSample", (event) => {
+      sampleOrigin = event.detail.origin;
+      void event.detail.respond({
+        model: "test-model",
+        stopReason: "endTurn",
+        role: "assistant",
+        content: { type: "text", text: "hello" },
+      });
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_sample");
+    const result = await connected.callTool(tool!, {});
+    expect(result.success).toBe(true);
+    expect(sampleOrigin).toBe("input-required");
+    const content = result.result!.content as ContentBlock[];
+    const text = content[0] && "text" in content[0] ? content[0].text : "";
+    expect(text).toContain("MRTR sample complete");
+  });
+
+  it("echoes a declined elicitation back to the server (decline is not an abort)", async () => {
+    const started = await startMrtrServer(createMrtrTool());
+    const connected = await connectWithEra(started.url, "modern");
+
+    connected.addEventListener("newPendingElicitation", (event) => {
+      void event.detail.respond({ action: "decline" });
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_confirm");
+    // The tool still completes — the declined result is echoed back, and the
+    // server returns its final result citing the declined answer.
+    const result = await connected.callTool(tool!, { action: "deploy" });
+    expect(result.success).toBe(true);
+    const content = result.result!.content as ContentBlock[];
+    const text = content[0] && "text" in content[0] ? content[0].text : "";
+    expect(text).toContain("MRTR complete");
+    expect(text).toContain("decline");
+  });
+
+  it("bounds a pathological MRTR that never completes (MRTR_MAX_ROUNDS)", async () => {
+    const started = await startMrtrServer(createMrtrLoopTool());
+    const connected = await connectWithEra(started.url, "modern");
+
+    // Auto-answer every round; the server never completes, so the driver's cap
+    // must stop the loop rather than spin forever.
+    connected.addEventListener("newPendingElicitation", (event) => {
+      void event.detail.respond({ action: "accept", content: { value: "x" } });
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_loop");
+    await expect(connected.callTool(tool!, {})).rejects.toThrow(
+      /exceeded .* input_required rounds/,
+    );
+  });
+
+  it("handles requestState-only and inputRequests-only rounds (param shaping)", async () => {
+    const started = await startMrtrServer(createMrtrEdgeCaseTool());
+    const connected = await connectWithEra(started.url, "modern");
+    const toolCallFrames = collectToolCallRequests(connected);
+
+    connected.addEventListener("newPendingElicitation", (event) => {
+      void event.detail.respond({ action: "accept", content: { value: "n" } });
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_edge");
+    const result = await connected.callTool(tool!, {});
+    expect(result.success).toBe(true);
+    const content = result.result!.content as ContentBlock[];
+    const text = content[0] && "text" in content[0] ? content[0].text : "";
+    expect(text).toContain("MRTR edge complete");
+
+    // Three rounds: original + two retries.
+    expect(toolCallFrames.length).toBe(3);
+    // Round-1 retry carries inputResponses but NO requestState (round 1 minted
+    // none).
+    const retry1 = toolCallFrames[1].params as {
+      inputResponses?: unknown;
+      requestState?: unknown;
+    };
+    expect(retry1.inputResponses).toBeDefined();
+    expect(retry1.requestState).toBeUndefined();
+    // Round-2 retry carries requestState and no meaningful inputResponses (round
+    // 2 was requestState-only; the driver adds none — the modern SDK codec may
+    // serialize an empty `{}` on the wire).
+    const retry2 = toolCallFrames[2].params as {
+      inputResponses?: Record<string, unknown>;
+      requestState?: unknown;
+    };
+    expect(retry2.requestState).toBeDefined();
+    expect(Object.keys(retry2.inputResponses ?? {})).toHaveLength(0);
+  });
+
+  it("cancels an in-flight MRTR call while its embedded request is pending", async () => {
+    const started = await startMrtrServer(createMrtrTool());
+    const connected = await connectWithEra(started.url, "modern");
+
+    // Cancel the tool call the moment its embedded elicitation is surfaced,
+    // instead of answering it.
+    connected.addEventListener("newPendingElicitation", () => {
+      connected.cancelToolCall();
+    });
+
+    const { tools } = await connected.listTools();
+    const tool = tools.find((t) => t.name === "mrtr_confirm");
+    await expect(
+      connected.callTool(tool!, { action: "deploy" }),
+    ).rejects.toThrow(ToolCallCancelledError);
+    // The pending elicitation was removed when the call aborted.
+    expect(connected.getPendingElicitations()).toHaveLength(0);
   });
 
   it("rejects a legacy client against a strict modern-only server", async () => {

core/mcp/elicitationCreateMessage.ts

@@ -1,5 +1,6 @@
 import type { ElicitRequest, ElicitResult } from "@modelcontextprotocol/client";
 import { RELATED_TASK_META_KEY } from "@modelcontextprotocol/client";
+import type { PendingRequestOrigin } from "./types.js";
 
 export type { ElicitRequest, ElicitResult };
 
@@ -13,6 +14,7 @@ export interface InspectorPendingElicitation {
   timestamp: Date;
   request: ElicitRequest;
   taskId?: string;
+  origin: PendingRequestOrigin;
 }
 
 /**
@@ -23,8 +25,19 @@ export class ElicitationCreateMessage {
   public readonly timestamp: Date;
   public readonly request: ElicitRequest;
   public readonly taskId?: string;
+  /**
+   * How this request reached the Inspector — a legacy server→client request or
+   * a modern MRTR `input_required` round. Drives era-accurate copy in the
+   * pending-request UI. Defaults to `"server-request"` so existing call sites
+   * (and stories) keep the legacy semantics unchanged.
+   */
+  public readonly origin: PendingRequestOrigin;
   private resolvePromise?: (result: ElicitResult) => void;
-  /** Set only for task-augmented elicit; used when user declines so server's tasks/result receives an error */
+  /**
+   * Rejects the originating call with an error. Set for task-augmented elicit
+   * (so the server's `tasks/result` receives the error on decline) and for
+   * MRTR-driven elicitations (so a genuine failure aborts the tool call).
+   */
   private rejectCallback?: (error: Error) => void;
   private onRemove: (id: string) => void;
 
@@ -33,6 +46,7 @@ export class ElicitationCreateMessage {
     resolve: (result: ElicitResult) => void,
     onRemove: (id: string) => void,
     reject?: (error: Error) => void,
+    origin: PendingRequestOrigin = "server-request",
   ) {
     this.onRemove = onRemove;
     this.id = `elicitation-${crypto.randomUUID()}`;
@@ -41,6 +55,7 @@ export class ElicitationCreateMessage {
     // Extract taskId from request params metadata if present
     const relatedTask = request.params?._meta?.[RELATED_TASK_META_KEY];
     this.taskId = relatedTask?.taskId;
+    this.origin = origin;
     this.resolvePromise = resolve;
     this.rejectCallback = reject;
   }

core/mcp/inspectorClient.ts

@@ -14,6 +14,7 @@ import type {
   ToolCallInvocation,
   AppRendererClient,
   InspectorClientOptions,
+  PendingRequestOrigin,
 } from "./types.js";
 // Re-export so v1.5 tests that do `import { InspectorClientOptions } from
 // "@inspector/core/mcp/inspectorClient.js"` keep resolving.
@@ -78,6 +79,7 @@ import type {
   ReadResourceRequest,
   GetPromptRequest,
   CompleteRequest,
+  ListRootsRequest,
 } from "@modelcontextprotocol/client";
 import type { Transport } from "@modelcontextprotocol/client";
 import type {
@@ -86,11 +88,20 @@ import type {
   VersionNegotiationOptions,
   ProtocolEra,
   DiscoverResult,
+  InputRequests,
+  InputRequiredOptions,
+  StandardSchemaV1,
 } from "@modelcontextprotocol/client";
 import { ProtocolError, ProtocolErrorCode } from "@modelcontextprotocol/client";
+import {
+  isInputRequiredResult,
+  withInputRequired,
+} from "@modelcontextprotocol/client";
 import {
   EmptyResultSchema,
   CallToolResultSchema,
+  GetPromptResultSchema,
+  ReadResourceResultSchema,
   // Task request schemas — used for `.shape.params` in the 3-arg custom
   // `setRequestHandler` form (tasks/* are excluded from v2's spec-method set).
   ListTasksRequestSchema,
@@ -173,6 +184,17 @@ interface ReceiverTaskRecord {
  */
 const MAX_URL_ELICITATION_RETRIES = 5;
 
+/**
+ * Error used to reject a pending sampling/elicitation request when the tool
+ * call driving its MRTR round is aborted (e.g. the user hits Cancel). Its
+ * message is not surfaced directly — `callToolWithRetries` maps the abort to a
+ * {@link ToolCallCancelledError} by inspecting the controller's reason — but a
+ * concrete error is needed to reject the awaiting driver promise.
+ */
+function createPendingAbortError(): Error {
+  return new Error("Pending request aborted");
+}
+
 /**
  * The abort reason used by `cancelToolCall()`. It rides along on the
  * `notifications/cancelled` sent to the server and lets `callToolWithRetries`
@@ -229,6 +251,14 @@ interface ToolCallRequest {
  * - Access to client functionality (prompts, resources, tools)
  */
 export class InspectorClient extends InspectorClientEventTarget {
+  /**
+   * Upper bound on MRTR (`input_required`) rounds for a single logical request
+   * before {@link requestWithInputRequired} gives up. We drive the loop
+   * ourselves (`inputRequired: { autoFulfill: false }`), so this is the manual
+   * counterpart to the SDK auto-driver's default `maxRounds` (10) and guards
+   * against a server that keeps returning `input_required` forever.
+   */
+  private static readonly MRTR_MAX_ROUNDS = 10;
   private client: Client | null = null;
   private appRendererClientProxy: AppRendererClient | null = null;
   // Lazily-built validator used only on the skipOutputValidation path to detect
@@ -403,12 +433,21 @@ export class InspectorClient extends InspectorClientEventTarget {
     const clientOptions: {
       capabilities?: ClientCapabilities;
       versionNegotiation?: VersionNegotiationOptions;
+      inputRequired?: InputRequiredOptions;
     } = {
       // Per-server protocol era (SEP §7.8), threaded from config via
       // `eraToVersionNegotiation` and defaulted to `{ mode: "legacy" }` in the
       // constructor. "legacy" keeps the wire byte-identical to a 2025 client;
       // "auto"/"modern" opt into 2026-era negotiation (#1626).
       versionNegotiation: this.versionNegotiation,
+      // Drive MRTR (SEP-2322) manually instead of letting the SDK auto-fulfil
+      // and hide the retry loop (#1704). Unconditional and safe on every era:
+      // legacy servers never return `input_required`, so this is a no-op there;
+      // on modern connections the three multi-round-trip methods opt in via
+      // `allowInputRequired` in `requestWithInputRequired`, and no other method
+      // can receive an `input_required` result. The negotiated era is unknown
+      // at construction time, so gating here is impossible anyway.
+      inputRequired: { autoFulfill: false },
     };
     const capabilities: ClientCapabilities = {};
     if (this.sample) {
@@ -1078,19 +1117,7 @@ export class InspectorClient extends InspectorClientEventTarget {
               taskResult as unknown as CreateMessageResult,
             );
           }
-          return new Promise<CreateMessageResult>((resolve, reject) => {
+          return this.enqueuePendingSample(request, "server-request");
-            const samplingRequest = new SamplingCreateMessage(
-              request,
-              (result) => {
-                resolve(result);
-              },
-              (error) => {
-                reject(error);
-              },
-              (id) => this.removePendingSample(id),
-            );
-            this.addPendingSample(samplingRequest);
-          });
         };
         this.client.setRequestHandler(
           "sampling/createMessage",
@@ -1158,16 +1185,7 @@ export class InspectorClient extends InspectorClientEventTarget {
             const taskResult: CreateTaskResult = { task: record.task };
             return Promise.resolve(taskResult as unknown as ElicitResult);
           }
-          return new Promise<ElicitResult>((resolve) => {
+          return this.enqueuePendingElicitation(request, "server-request");
-            const elicitationRequest = new ElicitationCreateMessage(
-              request,
-              (result) => {
-                resolve(result);
-              },
-              (id) => this.removePendingElicitation(id),
-            );
-            this.addPendingElicitation(elicitationRequest);
-          });
         };
         this.client.setRequestHandler("elicitation/create", elicitHandler);
         if (this.receiverTasks) {
@@ -1683,6 +1701,217 @@ export class InspectorClient extends InspectorClientEventTarget {
     return { tasks: result.tasks as Task[], nextCursor: result.nextCursor };
   }
 
+  /**
+   * Surface a sampling request through the pending-request UI and resolve with
+   * the user's answer. Shared by the inbound `sampling/createMessage` handler
+   * (legacy server→client request) and the MRTR driver (a modern
+   * `input_required` round embeds the request in a tool-call result). `origin`
+   * tags which of the two so the UI can show era-accurate semantics.
+   *
+   * Sampling has no decline/cancel action (unlike elicitation): the panel
+   * either sends a `CreateMessageResult` (resolve — echoed to the server) or
+   * Rejects (reject — which fails the tool call). A `signal` abort likewise
+   * rejects so the MRTR driver can abort the originating call. `signal` is only
+   * passed by the MRTR driver (the tool call's abort signal) — while the driver
+   * awaits an answer there is no in-flight SDK request to carry it.
+   */
+  private enqueuePendingSample(
+    request: CreateMessageRequest,
+    origin: PendingRequestOrigin,
+    signal?: AbortSignal,
+  ): Promise<CreateMessageResult> {
+    // A Promise's resolve/reject is idempotent (first settle wins, later ones
+    // no-op), so the respond/reject/abort paths need no extra guard against a
+    // double settle.
+    return new Promise<CreateMessageResult>((resolvePromise, rejectPromise) => {
+      const sample = new SamplingCreateMessage(
+        request,
+        resolvePromise,
+        rejectPromise,
+        (id) => this.removePendingSample(id),
+        origin,
+      );
+      this.addPendingSample(sample);
+      this.wirePendingAbort(signal, () => {
+        this.removePendingSample(sample.id);
+        rejectPromise(createPendingAbortError());
+      });
+    });
+  }
+
+  /**
+   * Surface an elicitation request through the pending-request UI and resolve
+   * with the user's answer. Shared by the inbound `elicitation/create` handler
+   * and the MRTR driver — see {@link enqueuePendingSample} for the `origin` /
+   * `signal` semantics. A declined/cancelled elicitation resolves with the
+   * corresponding `ElicitResult` (echoed to the server on retry); only a
+   * genuine failure or a `signal` abort rejects.
+   */
+  private enqueuePendingElicitation(
+    request: ElicitRequest,
+    origin: PendingRequestOrigin,
+    signal?: AbortSignal,
+  ): Promise<ElicitResult> {
+    // See {@link enqueuePendingSample} — Promise settle is idempotent.
+    return new Promise<ElicitResult>((resolvePromise, rejectPromise) => {
+      const elicitation = new ElicitationCreateMessage(
+        request,
+        resolvePromise,
+        (id) => this.removePendingElicitation(id),
+        rejectPromise,
+        origin,
+      );
+      this.addPendingElicitation(elicitation);
+      this.wirePendingAbort(signal, () => {
+        this.removePendingElicitation(elicitation.id);
+        rejectPromise(createPendingAbortError());
+      });
+    });
+  }
+
+  /**
+   * Reject a still-pending request when `signal` aborts (e.g. the user cancels
+   * the tool call while its MRTR round is awaiting an answer). No-op when
+   * `signal` is absent — the legacy inbound-handler path passes none.
+   */
+  private wirePendingAbort(
+    signal: AbortSignal | undefined,
+    onAbort: () => void,
+  ): void {
+    if (!signal) return;
+    /* v8 ignore next 4 -- unreachable in the MRTR flow: an abort during the
+       SDK request leg rejects `client.request` before we reach the pending
+       enqueue, so the signal is never already-aborted here; kept as a defensive
+       guard because addEventListener("abort") would not fire on a pre-aborted
+       signal. */
+    if (signal.aborted) {
+      onAbort();
+      return;
+    }
+    signal.addEventListener("abort", onAbort, { once: true });
+  }
+
+  /**
+   * Drive a multi-round-trip request (SEP-2322 "MRTR") for one of the modern
+   * multi-round-trip methods (`tools/call`, `prompts/get`, `resources/read`).
+   *
+   * The client is constructed with `inputRequired: { autoFulfill: false }`, so
+   * an `input_required` result is handed back here (via `allowInputRequired`)
+   * instead of the SDK silently fulfilling and retrying. We surface each
+   * embedded request through the SAME pending-request UI the legacy
+   * server→client path uses (`fulfilInputRequests`), gather the bare results,
+   * and retry the ORIGINAL request with `inputResponses` + the echoed
+   * `requestState` on a fresh JSON-RPC id (`client.request` mints it). The loop
+   * runs until the server returns a complete result, bounded by
+   * {@link MRTR_MAX_ROUNDS}.
+   *
+   * On legacy connections a server never returns `input_required`, so the first
+   * response is always complete and this is a single `client.request` call.
+   */
+  private async requestWithInputRequired<TSchema extends StandardSchemaV1>(
+    method: "tools/call" | "prompts/get" | "resources/read",
+    params: Record<string, unknown>,
+    resultSchema: TSchema,
+    requestOptions: RequestOptions,
+  ): Promise<StandardSchemaV1.InferOutput<TSchema>> {
+    const client = this.client;
+    /* v8 ignore next 3 -- defensive: every caller (callTool/getPrompt/
+       readResource) already verified this.client is non-null before reaching
+       here, so this guard cannot trip in practice. */
+    if (!client) {
+      throw new Error("Client is not connected");
+    }
+    const wrapped = withInputRequired(resultSchema);
+    const signal = requestOptions.signal;
+    let round = 0;
+    let nextParams = params;
+    while (true) {
+      const outcome = await client.request(
+        { method, params: nextParams },
+        wrapped,
+        {
+          ...requestOptions,
+          allowInputRequired: true,
+        },
+      );
+      if (!isInputRequiredResult(outcome)) {
+        return outcome;
+      }
+      round += 1;
+      if (round > InspectorClient.MRTR_MAX_ROUNDS) {
+        throw new Error(
+          `Multi-round-trip "${method}" exceeded ${InspectorClient.MRTR_MAX_ROUNDS} input_required rounds without completing.`,
+        );
+      }
+      const inputResponses = await this.fulfilInputRequests(
+        outcome.inputRequests,
+        signal,
+      );
+      // Retry re-issues the ORIGINAL params plus THIS round's answers and the
+      // server's opaque state token; the SDK assigns a fresh JSON-RPC id.
+      nextParams = {
+        ...params,
+        ...(inputResponses ? { inputResponses } : {}),
+        ...(outcome.requestState !== undefined
+          ? { requestState: outcome.requestState }
+          : {}),
+      };
+    }
+  }
+
+  /**
+   * Fulfil the embedded requests of one MRTR `input_required` round, keyed by
+   * the server-assigned identifiers echoed back in `inputResponses`. Sequential
+   * (one modal at a time) to keep the single-slot pending UI coherent. Returns
+   * `undefined` for a `requestState`-only round (no embedded requests); an empty
+   * `inputRequests` map yields an empty `{}`, which the retry echoes harmlessly.
+   */
+  private async fulfilInputRequests(
+    inputRequests: InputRequests | undefined,
+    signal?: AbortSignal,
+  ): Promise<Record<string, unknown> | undefined> {
+    if (!inputRequests) return undefined;
+    const responses: Record<string, unknown> = {};
+    for (const [key, embedded] of Object.entries(inputRequests)) {
+      responses[key] = await this.fulfilEmbeddedInputRequest(embedded, signal);
+    }
+    return responses;
+  }
+
+  /**
+   * Fulfil a single embedded MRTR request. `roots/list` is auto-answered from
+   * the configured roots (consistent with the legacy `roots/list` handler — no
+   * pending UX); `elicitation/create` and `sampling/createMessage` surface
+   * through the pending-request UI tagged `"input-required"`.
+   */
+  private async fulfilEmbeddedInputRequest(
+    request: CreateMessageRequest | ElicitRequest | ListRootsRequest,
+    signal?: AbortSignal,
+  ): Promise<unknown> {
+    switch (request.method) {
+      case "roots/list":
+        return { roots: this.roots ?? [] };
+      case "elicitation/create":
+        return this.enqueuePendingElicitation(
+          request,
+          "input-required",
+          signal,
+        );
+      case "sampling/createMessage":
+        return this.enqueuePendingSample(request, "input-required", signal);
+      /* v8 ignore next 6 -- defensive: an SDK server rejects an unknown embedded
+         method before it reaches the wire, so this only guards against a
+         non-conformant hand-rolled server; not reproducible against the
+         SDK-based test servers. */
+      default:
+        throw new Error(
+          `Unsupported embedded input_required request method: ${
+            (request as { method: string }).method
+          }`,
+        );
+    }
+  }
+
   /**
    * Get all pending sampling requests
    */
@@ -1951,10 +2180,16 @@ export class InspectorClient extends InspectorClientEventTarget {
         return await this.attemptToolCall(request, abortController.signal);
       } catch (error) {
         // The controller was aborted. A deliberate `cancelToolCall()` (matched
-        // by reason) means the SDK already sent `notifications/cancelled` — so
+        // by reason) means the SDK already sent `notifications/cancelled` if the
-        // surface a clean cancellation, not a generic failure, and don't record
+        // abort landed during a `client.request` leg — so surface a clean
-        // it in history. Any other abort (e.g. a disconnect, which aborts with a
+        // cancellation, not a generic failure, and don't record it in history.
-        // different reason) falls through to the normal error path (#1458).
+        // If instead the abort lands while an MRTR round is awaiting an embedded
+        // pending request (between `client.request` legs), there is no in-flight
+        // SDK request, so nothing is sent on the wire — `wirePendingAbort` just
+        // rejects the pending request and the driver abandons the retry; the
+        // outcome here is identical. Any other abort (e.g. a disconnect, which
+        // aborts with a different reason) falls through to the normal error path
+        // (#1458).
         if (
           abortController.signal.aborted &&
           abortController.signal.reason === TOOL_CALL_CANCELLED_REASON
@@ -2092,29 +2327,28 @@ export class InspectorClient extends InspectorClientEventTarget {
       callParams.task = { ttl: taskOptions.ttl };
     }
 
-    // MCP Apps forward the server's CallToolResult straight to the running
-    // view, which is the real consumer. The SDK's callTool() validates
-    // structuredContent against the tool's outputSchema and THROWS on a
-    // mismatch — which would deny the app a result the server actually
-    // returned (and that legacy hosts render fine). For those passthrough
-    // calls go through request() directly, which skips that host-side
-    // validation. Regular Tools-screen calls keep validating.
     const requestOptions = this.getRequestOptions(
       metadata?.progressToken,
       signal,
     );
-    // Both branches yield a CallToolResult: request() parsed it with
+    // Route through the MRTR driver (`requestWithInputRequired`) so a modern
-    // CallToolResultSchema above, callTool() returns the same shape — so the
+    // `input_required` result pauses at the pending-request UI and retries with
-    // `as CallToolResult` casts below are safe.
+    // the user's answer (#1704). Both eras use `client.request` with
+    // `CallToolResultSchema`; on legacy this is a single round. We deliberately
+    // do NOT use `client.callTool` (which would auto-fulfil / reject on an
+    // `input_required` result) — its only extra behavior over `request` is
+    // structuredContent output validation, which we already re-implement below
+    // via `validateToolOutput`. MCP Apps passthrough (skipOutputValidation)
+    // simply skips that check; both paths yield a CallToolResult once the
+    // driver returns a complete (non-`input_required`) result.
     const result = await this.invokeMcpClient(
       () =>
-        options?.skipOutputValidation
+        this.requestWithInputRequired(
-          ? client.request(
+          "tools/call",
-              { method: "tools/call", params: callParams },
+          callParams,
-              CallToolResultSchema,
+          CallToolResultSchema,
-              requestOptions,
+          requestOptions,
-            )
+        ),
-          : client.callTool(callParams, requestOptions),
       { method: "tools/call", toolName: tool.name },
     );
 
@@ -2126,10 +2360,7 @@ export class InspectorClient extends InspectorClientEventTarget {
     //    what a strict host would do), so the caller sees the error.
     //  - skipOutputValidation (MCP Apps passthrough): non-fatal — surface it as
     //    an advisory so a schema-violating-but-real result still reaches the app.
-    const outputValidationError = this.validateToolOutput(
+    const outputValidationError = this.validateToolOutput(tool, result);
-      tool,
-      result as CallToolResult,
-    );
     if (outputValidationError && !options?.skipOutputValidation) {
       // Match the prior contract: on v1 a strict output-schema violation
       // surfaced as the SDK's typed `McpError`/`ProtocolError` (code
@@ -2144,7 +2375,7 @@ export class InspectorClient extends InspectorClientEventTarget {
     const invocation: ToolCallInvocation = {
       toolName: tool.name,
       params: args,
-      result: result as CallToolResult,
+      result,
       timestamp,
       success: true,
       metadata,
@@ -2295,6 +2526,10 @@ export class InspectorClient extends InspectorClientEventTarget {
     // synchronously (or for which the tool forbids/ignores task augmentation)
     // may return an immediate `CallToolResult` instead — accept either with a
     // union schema and branch on the presence of `task`.
+    //
+    // NOTE: this task path does NOT opt into `allowInputRequired`, so a
+    // task-augmented tool that returns `input_required` is not MRTR-driven here.
+    // Driving MRTR over the tasks extension is out of scope for #1704.
     const requestPromise = client.request(
       { method: "tools/call", params },
       CreateTaskResultSchema.or(CallToolResultSchema),
@@ -2669,10 +2904,15 @@ export class InspectorClient extends InspectorClientEventTarget {
       uri,
       ...(effectiveMeta ? { _meta: effectiveMeta } : {}),
     };
+    // MRTR-driven (#1704): a modern `resources/read` can return `input_required`
+    // (embedding an elicitation/sampling request); the driver pauses at the
+    // pending-request UI and retries with the answer. Legacy is a single round.
     const result = await this.invokeMcpClient(
       () =>
-        this.client!.readResource(
+        this.requestWithInputRequired(
+          "resources/read",
           params,
+          ReadResourceResultSchema,
           this.getRequestOptions(metadata?.progressToken),
         ),
       { method: "resources/read" },
@@ -2836,10 +3076,15 @@ export class InspectorClient extends InspectorClientEventTarget {
       ...(effectiveMeta ? { _meta: effectiveMeta } : {}),
     };
 
+    // MRTR-driven (#1704): a modern `prompts/get` can return `input_required`;
+    // the driver pauses at the pending-request UI and retries with the answer.
+    // Legacy is a single round.
     const result = await this.invokeMcpClient(
       () =>
-        this.client!.getPrompt(
+        this.requestWithInputRequired(
+          "prompts/get",
           params,
+          GetPromptResultSchema,
           this.getRequestOptions(metadata?.progressToken),
         ),
       { method: "prompts/get", toolName: name },