diff --git a/packages/paperclip-runner/README.md b/packages/paperclip-runner/README.md
index 7c0647c3094..7f03e6c14e2 100644
--- a/packages/paperclip-runner/README.md
+++ b/packages/paperclip-runner/README.md
@@ -16,9 +16,13 @@ and a real-model example against the mock core. These phases do not import or
 start Paperclip's server, UI, CLI, or production database. The Phase 4b lower
 layer adds browser-resolved provider requests, goal capability detection,
 subagent lineage, control-race semantics, and a package-local demo server that
-keeps Codex authentication server-side.
+keeps Codex authentication server-side. The Phase 4b live console adds the
+browser tracer over that boundary: a live transcript and composer, same-turn
+steering, three distinct interrupt races, inline request cards,
+capability-gated goal controls, parent/child lineage, reconnect, refresh,
+replay, and a redacted protocol inspector.
 
-## Phase 0–4 quick start
+## Phase 0–4b quick start
 
 From the repository root:
 
@@ -59,6 +63,9 @@ pnpm --filter @paperclipai/paperclip-runner trace:phase2 -- --scenario happy-pat
 pnpm --filter @paperclipai/paperclip-runner trace:phase3 -- --fault lost-ack
 pnpm --filter @paperclipai/paperclip-runner trace:phase4
 pnpm --filter @paperclipai/paperclip-runner demo:phase4b -- --host 127.0.0.1 --port 4174
+
+# Phase 4b: chat with a live session in the browser.
+pnpm --filter @paperclipai/paperclip-runner console:phase4b
 pnpm --filter @paperclipai/paperclip-runner browser:dev --host 127.0.0.1 --port 4179
 ```
 
@@ -85,10 +92,11 @@ pnpm --filter @paperclipai/paperclip-runner browser:dev --host 127.0.0.1 --port
 | `trace:phase4` | Run the mock core with a real, local skillless Codex app-server session. |
 | `record:phase4` | Run the safe Codex task and record its validated, normalized trace. |
 | `demo:phase4b` | Start the package-local HTTP/SSE server with server-only Codex authentication. |
+| `console:phase4b` | Start the standalone browser devtool with the Phase 4b live console on `127.0.0.1:4180`. |
 | `record:phase4b` | Run a safe real Codex task through the demo server and record reconnect/replay evidence. |
 | `browser:dev` | Start the standalone live/replay/recovery browser devtool. |
 | `test:browser` | Exercise static replay and live scenarios, then capture temporary screenshots under ignored test output. |
-| `verify` | Run the complete deterministic Phase 0 through Phase 4 acceptance sequence. |
+| `verify` | Run the complete deterministic Phase 0 through Phase 4b acceptance sequence. |
 | `verify:rootless` | Extract Debian/Ubuntu browser libraries without root, then run `verify`. |
 
 ## Navigate
@@ -105,6 +113,8 @@ pnpm --filter @paperclipai/paperclip-runner browser:dev --host 127.0.0.1 --port
 - [Phase 4 skillless Codex driver reference](docs/phase-04-skillless-codex-driver.md)
 - [Phase 4b protocol/server tutorial](docs/tutorials/phase-04b-protocol-server.md)
 - [Phase 4b protocol/server reference](docs/phase-04b-protocol-server.md)
+- [Phase 4b live console tutorial](docs/tutorials/phase-04b-live-console.md)
+- [Phase 4b live console reference](docs/phase-04b-live-console.md)
 - [PRP compatibility/versioning policy](docs/protocol-compatibility.md)
 - [Cumulative end-to-end tutorial](docs/tutorials/end-to-end.md)
 - [Journal guide](docs/journal.md)
@@ -112,5 +122,6 @@ pnpm --filter @paperclipai/paperclip-runner browser:dev --host 127.0.0.1 --port
 - [Implementation plan](spec/paperclip-native-runner-implementation-plan.md)
 - [Normative spike specification](spec/paperclip-native-runner-spike-spec.md)
 
-Phase 4 adds the package-local real-model reference driver. Production
-Paperclip integration and Phase 5 remain deferred.
+Phase 4 adds the package-local real-model reference driver and Phase 4b adds
+the package-local browser console over it. Production Paperclip integration and
+Phase 5 remain deferred.
diff --git a/packages/paperclip-runner/devtools/browser/live-console.spec.ts b/packages/paperclip-runner/devtools/browser/live-console.spec.ts
new file mode 100644
index 00000000000..3a03fe77171
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/live-console.spec.ts
@@ -0,0 +1,395 @@
+import { resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+
+import { expect, test, type Page } from "@playwright/test";
+
+const evidenceRoot = resolve(
+  fileURLToPath(new URL("../../knowledge/evidence/phase-04b", import.meta.url)),
+);
+
+const DESKTOP = { width: 1440, height: 900 };
+const MOBILE = { width: 390, height: 844 };
+
+function shot(page: Page, name: string) {
+  return page.screenshot({ path: resolve(evidenceRoot, `${name}.png`), fullPage: true });
+}
+
+async function openConsole(page: Page) {
+  await page.goto("/");
+  await page.getByRole("button", { name: "Live console" }).click();
+  await expect(page.getByRole("heading", { name: "Live Codex protocol console" })).toBeVisible();
+}
+
+async function runManifest(page: Page, id: string) {
+  await expect(page.getByTestId("manifest-list")).toBeVisible();
+  await page.getByTestId(`manifest-${id}`).click();
+  await page.getByTestId("run-manifest").click();
+  await expect(page.getByTestId("connection-status")).toHaveText("connected", { timeout: 15_000 });
+}
+
+test.describe("Phase 4b live console", () => {
+  test.beforeEach(async ({ page }) => {
+    await page.setViewportSize(DESKTOP);
+  });
+
+  test("shows the empty state and runs a completed turn", async ({ page }) => {
+    await openConsole(page);
+    await expect(page.getByTestId("empty-transcript")).toBeVisible();
+    await expect(page.getByTestId("composer-send")).toBeDisabled();
+    await shot(page, "desktop-01-idle-empty");
+
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("reasoning-item")).toBeVisible();
+    await expect(page.getByTestId("tool-item")).toBeVisible();
+    await shot(page, "desktop-02-streaming-turn");
+
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+    await expect(page.getByTestId("live-transcript")).toContainText("finished the task");
+    await expect(page.getByTestId("live-transcript")).toContainText(
+      "Summarize what the Phase 4b demo server proves.",
+    );
+    await expect(page.getByTestId("tool-item")).toHaveAttribute("data-status", "completed");
+    await shot(page, "desktop-03-turn-completed");
+
+    // The browser must never receive a provider credential.
+    const body = await page.locator("body").innerText();
+    expect(body).not.toContain("Bearer ");
+    expect(body).not.toContain("auth.json");
+  });
+
+  test("acknowledges steering and keeps rejected text recoverable", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "steering");
+    await expect(page.getByTestId("composer-send")).toHaveText("Steer", { timeout: 15_000 });
+
+    await page.getByTestId("composer-input").fill("Keep it to one sentence.");
+    await page.getByTestId("composer-send").click();
+    const chip = page.getByTestId("steering-chip").first();
+    await expect(chip).toContainText("acknowledged", { timeout: 15_000 });
+    await shot(page, "desktop-04-steering-acknowledged");
+
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 25_000 });
+    // The composer state machine now offers Send again; force a stale steer.
+    await page.evaluate(() => window.history.replaceState({}, "", "/"));
+  });
+
+  test("rejects a stale steer and offers to resend it", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "steering");
+    await expect(page.getByTestId("composer-send")).toHaveText("Steer", { timeout: 15_000 });
+    await page.getByTestId("composer-input").fill("Typed while the turn was running.");
+
+    // Let the turn finish before the steer is sent: the classic stale race.
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 25_000 });
+    await expect(page.getByTestId("composer-send")).toHaveText("Send");
+  });
+
+  test("cancels a turn interrupted before it starts", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "interrupt-before-start");
+    await expect(page.getByTestId("composer-stop")).toBeEnabled();
+    await page.getByTestId("composer-stop").click();
+    await expect(page.getByTestId("turn-cancelled")).toBeVisible({ timeout: 15_000 });
+    await expect(page.getByTestId("turn-cancelled")).toContainText("Cancelled before start");
+    await shot(page, "desktop-05-interrupt-before-start");
+  });
+
+  test("keeps partial text when a streaming turn is stopped", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "interrupt-during-generation");
+    await expect(page.getByTestId("live-transcript")).toContainText("The tracer keeps", {
+      timeout: 15_000,
+    });
+    await page.getByTestId("composer-stop").click();
+    await expect(page.getByTestId("turn-interrupted")).toBeVisible({ timeout: 15_000 });
+    await expect(page.getByTestId("interrupted-divider").first()).toBeVisible();
+    await expect(page.getByTestId("live-transcript")).not.toContainText("And this final one.");
+    await shot(page, "desktop-06-interrupt-during-generation");
+  });
+
+  test("keeps an interrupted tool item without a fabricated result", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "interrupt-during-tool");
+    await expect(page.getByTestId("tool-item")).toHaveAttribute("data-status", "running", {
+      timeout: 15_000,
+    });
+    await page.getByTestId("composer-stop").click();
+    await expect(page.getByTestId("tool-item")).toHaveAttribute("data-status", "interrupted", {
+      timeout: 15_000,
+    });
+    await expect(page.getByTestId("turn-interrupted")).toBeVisible();
+    await shot(page, "desktop-07-interrupt-during-tool");
+  });
+
+  test("resolves approvals exactly once and collapses the resolved card", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "approvals");
+    const card = page.getByTestId("request-card").first();
+    await expect(card).toHaveAttribute("data-status", "pending", { timeout: 15_000 });
+    await expect(page.getByTestId("pending-request-banner")).toBeVisible();
+    await expect(page.getByTestId("request-action-accept_for_session")).toBeVisible();
+    await shot(page, "desktop-08-request-pending");
+
+    await page.getByTestId("request-action-accept").first().click();
+    await expect(page.getByTestId("request-card").first()).toHaveAttribute(
+      "data-status",
+      "resolved",
+      { timeout: 15_000 },
+    );
+    await expect(page.getByTestId("request-card").first()).toContainText("resolved — Approve");
+    await shot(page, "desktop-09-request-resolved");
+
+    // Second request only offers accept and decline.
+    const second = page.getByTestId("request-card").nth(1);
+    await expect(second).toHaveAttribute("data-status", "pending", { timeout: 15_000 });
+    await expect(second.getByTestId("request-action-accept_for_session")).toHaveCount(0);
+  });
+
+  test("submits a user-input answer and lets an unanswered request expire", async ({ page }) => {
+    test.setTimeout(60_000);
+    await openConsole(page);
+    await runManifest(page, "user-input");
+    await expect(page.getByTestId("request-answer")).toBeVisible({ timeout: 15_000 });
+    await page.getByTestId("request-answer").fill("staging");
+    await page.getByTestId("request-action-submit").click();
+    await expect(page.getByTestId("request-card").first()).toHaveAttribute(
+      "data-status",
+      "resolved",
+      { timeout: 15_000 },
+    );
+
+    await expect(page.getByTestId("request-card").nth(1)).toHaveAttribute(
+      "data-status",
+      "expired",
+      { timeout: 25_000 },
+    );
+    await expect(page.getByTestId("request-card").nth(1)).toContainText("expired before response");
+    await shot(page, "desktop-10-request-expired");
+  });
+
+  test("moves a goal through its lifecycle", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "goals");
+    await expect(page.getByTestId("goal-banner")).toHaveText("No goal set", { timeout: 15_000 });
+
+    await page.locator("#goal-menu").click();
+    await page.getByRole("menuitem", { name: "Set goal…" }).click();
+    await page.getByTestId("goal-objective").fill("Keep the demo goal visible");
+    await page.getByTestId("goal-set-confirm").click();
+    await expect(page.getByTestId("goal-banner")).toContainText("Keep the demo goal visible", {
+      timeout: 15_000,
+    });
+    await expect(page.getByTestId("goal-banner")).toContainText("active");
+
+    await page.locator("#goal-menu").click();
+    await page.getByRole("menuitem", { name: "Pause" }).click();
+    await expect(page.getByTestId("goal-banner")).toContainText("paused", { timeout: 15_000 });
+    await shot(page, "desktop-11-goal-supported");
+  });
+
+  test("disables the goal menu with the exact upstream diagnostic", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "goals-unsupported");
+    const trigger = page.locator("#goal-menu");
+    await expect(trigger).toHaveAttribute("aria-disabled", "true", { timeout: 15_000 });
+    const describedBy = await trigger.getAttribute("aria-describedby");
+    expect(describedBy).not.toBeNull();
+    await expect(page.locator(`#${describedBy}`)).toHaveText(
+      "Goal operations unsupported: app-server 0.132.0 does not advertise the goals capability",
+    );
+
+    await page.getByTestId("toggle-inspector").click();
+    await page.getByRole("tab", { name: "Capabilities" }).click();
+    await expect(page.getByTestId("inspector-capabilities")).toContainText(
+      "does not advertise the goals capability",
+    );
+    await shot(page, "desktop-12-goal-unsupported");
+  });
+
+  test("shows child threads and refuses to emulate child steering", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "subagents");
+    await expect(page.getByTestId("thread-thread-child-b")).toBeVisible({ timeout: 20_000 });
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+    await expect(page.getByTestId("thread-thread-child-a")).toContainText("Completed");
+    await expect(page.getByTestId("thread-thread-child-b")).toContainText("Failed");
+
+    await page.getByTestId("thread-thread-child-a").click();
+    await expect(page.getByTestId("breadcrumb")).toContainText("thread-child-a");
+    await expect(page.getByTestId("composer-reason")).toHaveText(
+      "Direct steering of child threads is not supported by this app-server.",
+    );
+    await expect(page.getByTestId("composer-input")).toBeDisabled();
+    await shot(page, "desktop-13-lineage-child");
+  });
+
+  test("shows the failed item and diagnostic in the record", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "failure");
+    await expect(page.getByTestId("turn-failed")).toBeVisible({ timeout: 20_000 });
+    await expect(page.getByTestId("tool-failure")).toContainText("cannot find module");
+    await shot(page, "desktop-14-failed-turn");
+  });
+
+  test("resumes the same session after a transport drop", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+    const before = await page.getByTestId("live-transcript").innerText();
+
+    await page.getByTestId("drop-connection").click();
+    await expect(page.getByTestId("reconnect-banner")).toBeVisible();
+    await expect(page.getByTestId("composer-input")).toBeDisabled();
+    await shot(page, "desktop-15-reconnect-banner");
+
+    await page.getByTestId("retry-now").click();
+    await expect(page.getByTestId("connection-status")).toHaveText("connected", {
+      timeout: 20_000,
+    });
+    // Live and resumed transcripts must be identical, not merely similar.
+    expect(await page.getByTestId("live-transcript").innerText()).toContain(before.trim());
+  });
+
+  test("restores the same session across a page refresh", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+    const before = await page.getByTestId("live-transcript").innerText();
+
+    await page.reload();
+    await page.getByRole("button", { name: "Live console" }).click();
+    await expect(page.getByTestId("live-transcript")).toContainText("finished the task", {
+      timeout: 20_000,
+    });
+    expect(await page.getByTestId("live-transcript").innerText()).toContain(before.trim());
+    await shot(page, "desktop-16-refresh-replay");
+  });
+
+  test("replays a recorded session with a stepper", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+
+    await page.getByTestId("toggle-replay").click();
+    await expect(page.getByTestId("replay-badge")).toBeVisible();
+    await expect(page.getByTestId("replay-stepper")).toBeVisible();
+    await expect(page.getByTestId("empty-transcript")).toBeVisible();
+
+    await page.getByRole("button", { name: "Step forward" }).click();
+    await page.getByRole("button", { name: "Step forward" }).click();
+    await expect(page.getByTestId("empty-transcript")).toHaveCount(0);
+    await shot(page, "desktop-17-replay-mode");
+
+    await page.getByTestId("toggle-replay").click();
+    await expect(page.getByTestId("replay-badge")).toHaveCount(0);
+    await expect(page.getByTestId("turn-completed")).toBeVisible();
+  });
+
+  test("inspects canonical events and reports reducer parity", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+
+    await page.getByTestId("toggle-inspector").click();
+    await expect(page.getByTestId("inspector-events")).toBeVisible();
+    await page.getByTestId("inspector-filter").fill("turn.completed");
+    await expect(page.getByTestId("inspector-events").getByRole("listitem")).toHaveCount(1);
+    await shot(page, "desktop-18-inspector-events");
+
+    await page.getByTestId("inspector-filter").fill("");
+    await page.getByRole("tab", { name: "Session" }).click();
+    await expect(page.getByTestId("inspector-session")).toContainText("Credentials in browser");
+    await expect(page.getByTestId("replay-parity")).toHaveText("match");
+    await shot(page, "desktop-19-inspector-session");
+  });
+
+  test("meets the keyboard contract", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "steering");
+    await expect(page.getByTestId("composer-send")).toHaveText("Steer", { timeout: 15_000 });
+
+    // K2: Escape moves focus to Stop and never interrupts directly.
+    await page.getByTestId("composer-input").focus();
+    await page.keyboard.press("Escape");
+    await expect(page.getByTestId("composer-stop")).toBeFocused();
+    await expect(page.getByTestId("turn-interrupted")).toHaveCount(0);
+
+    // K2: Enter steers the running turn.
+    await page.getByTestId("composer-input").fill("Steered with the keyboard.");
+    await page.getByTestId("composer-input").press("Enter");
+    await expect(page.getByTestId("steering-chip").first()).toContainText("acknowledged", {
+      timeout: 15_000,
+    });
+
+    // K7: inspector tabs follow the arrow-key tabs pattern.
+    await page.getByTestId("toggle-inspector").click();
+    await page.getByRole("tab", { name: "Events" }).focus();
+    await page.keyboard.press("ArrowRight");
+    await expect(page.getByRole("tab", { name: "Requests" })).toBeFocused();
+    await expect(page.getByRole("tab", { name: "Requests" })).toHaveAttribute(
+      "aria-selected",
+      "true",
+    );
+
+    // K5: lineage tree responds to arrow keys.
+    await page.getByTestId("lineage-tree").getByRole("treeitem").first().focus();
+    await page.keyboard.press("ArrowDown");
+
+    // K6/A1: the transcript is a labelled log landmark.
+    await expect(page.getByTestId("live-transcript")).toHaveAttribute("role", "log");
+    await expect(page.getByTestId("live-transcript")).toHaveAttribute("aria-live", "polite");
+  });
+
+  test("resets demo state behind a naming confirmation", async ({ page }) => {
+    await openConsole(page);
+    await runManifest(page, "completion");
+    await expect(page.getByTestId("turn-completed")).toBeVisible({ timeout: 20_000 });
+
+    await page.getByRole("button", { name: "Reset demo state", exact: true }).first().click();
+    await expect(page.getByRole("dialog")).toContainText(
+      "Recorded evidence files are not touched.",
+    );
+    await page.getByTestId("reset-confirm").click();
+    await expect(page.getByTestId("empty-transcript")).toBeVisible();
+  });
+});
+
+test.describe("Phase 4b live console on a small screen", () => {
+  test("keeps the transcript and composer usable at 390 by 844", async ({ page }) => {
+    await page.setViewportSize(MOBILE);
+    await openConsole(page);
+    await expect(page.getByRole("tab", { name: "Chat" })).toBeVisible();
+    await expect(page.getByTestId("manifest-list")).toHaveCount(0);
+    await shot(page, "mobile-01-idle-empty");
+
+    await page.getByRole("tab", { name: "Session" }).click();
+    await expect(page.getByTestId("manifest-list")).toBeVisible();
+    await page.getByTestId("manifest-approvals").click();
+    await page.getByTestId("run-manifest").click();
+    await shot(page, "mobile-02-session-segment");
+
+    await page.getByRole("tab", { name: "Chat" }).click();
+    await expect(page.getByTestId("request-card").first()).toHaveAttribute(
+      "data-status",
+      "pending",
+      { timeout: 20_000 },
+    );
+    await shot(page, "mobile-03-request-pending");
+    await page.getByTestId("request-action-accept").first().click();
+    await expect(page.getByTestId("request-card").first()).toHaveAttribute(
+      "data-status",
+      "resolved",
+      { timeout: 15_000 },
+    );
+
+    await page.getByRole("tab", { name: "Inspector" }).click();
+    await expect(page.getByTestId("inspector-events")).toBeVisible();
+    await shot(page, "mobile-04-inspector");
+
+    // WCAG 1.4.10 reflow: no horizontal scrolling of the page itself.
+    const overflow = await page.evaluate(
+      () => document.documentElement.scrollWidth - document.documentElement.clientWidth,
+    );
+    expect(overflow).toBeLessThanOrEqual(1);
+  });
+});
diff --git a/packages/paperclip-runner/devtools/browser/playwright.config.ts b/packages/paperclip-runner/devtools/browser/playwright.config.ts
index 0eb2bc3a5c6..217c591b688 100644
--- a/packages/paperclip-runner/devtools/browser/playwright.config.ts
+++ b/packages/paperclip-runner/devtools/browser/playwright.config.ts
@@ -2,7 +2,7 @@ import { defineConfig } from "@playwright/test";
 
 export default defineConfig({
   testDir: ".",
-  testMatch: "static-replay.spec.ts",
+  testMatch: ["static-replay.spec.ts", "live-console.spec.ts"],
   fullyParallel: false,
   workers: 1,
   reporter: "line",
@@ -13,6 +13,11 @@ export default defineConfig({
   webServer: {
     command:
       "pnpm exec vite preview --config vite.config.ts --host 127.0.0.1 --port 4179",
+    env: {
+      // Keep streamed chunks slow enough that interrupt races are reachable
+      // from a test, and fast enough that the suite stays quick.
+      PAPERCLIP_PHASE4B_CHUNK_DELAY_MS: "60",
+    },
     url: "http://127.0.0.1:4179",
     reuseExistingServer: false,
     timeout: 30_000,
diff --git a/packages/paperclip-runner/devtools/browser/src/App.tsx b/packages/paperclip-runner/devtools/browser/src/App.tsx
index c7d395043f4..695d93b022c 100644
--- a/packages/paperclip-runner/devtools/browser/src/App.tsx
+++ b/packages/paperclip-runner/devtools/browser/src/App.tsx
@@ -29,6 +29,7 @@ import {
   CardTitle,
 } from "./components/ui/card";
 import { Textarea } from "./components/ui/textarea";
+import { LiveConsole } from "./live/LiveConsole";
 
 const fixtureFiles = import.meta.glob(
   "../../../protocol/fixtures/phase-01/*.json",
@@ -53,7 +54,7 @@ const liveScenarios: Array<{ key: Phase2Scenario; label: string }> = [
   { key: "duplicate-terminal", label: "Duplicate terminal guard" },
 ];
 
-type BrowserMode = "recovery" | "live" | "replay";
+type BrowserMode = "console" | "recovery" | "live" | "replay";
 type LiveStatus = "idle" | "starting" | "running" | "terminal" | "error";
 
 const phase3FaultLabels: Record<Phase3Fault, string> = {
@@ -74,6 +75,12 @@ const modeCopy: Record<
   BrowserMode,
   { eyebrow: string; title: string; description: string }
 > = {
+  console: {
+    eyebrow: "Paperclip Runner Protocol · Phase 4b",
+    title: "Live Codex protocol console",
+    description:
+      "Chat with a live session, steer it, stop it, answer its requests, and inspect the canonical protocol behind every surface.",
+  },
   recovery: {
     eyebrow: "Paperclip Runner Protocol · Phase 3",
     title: "Durable recovery diagnostics",
@@ -821,6 +828,8 @@ function StaticReplay() {
 }
 
 export function App() {
+  // Phase 1-3 surfaces keep their landing mode: their screenshots are frozen
+  // QA evidence and their specs open on the Phase 2 runner.
   const [mode, setMode] = useState<BrowserMode>("live");
   const header = modeCopy[mode];
 
@@ -836,6 +845,13 @@ export function App() {
       </header>
 
       <nav className="mode-switch" aria-label="Runner devtool mode">
+        <Button
+          type="button"
+          aria-pressed={mode === "console"}
+          onClick={() => setMode("console")}
+        >
+          Live console
+        </Button>
         <Button
           type="button"
           aria-pressed={mode === "recovery"}
@@ -859,7 +875,15 @@ export function App() {
         </Button>
       </nav>
 
-      {mode === "recovery" ? <RecoveryDiagnostics /> : mode === "live" ? <LiveRunner /> : <StaticReplay />}
+      {mode === "console" ? (
+        <LiveConsole />
+      ) : mode === "recovery" ? (
+        <RecoveryDiagnostics />
+      ) : mode === "live" ? (
+        <LiveRunner />
+      ) : (
+        <StaticReplay />
+      )}
     </main>
   );
 }
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/badge.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/badge.tsx
index 261ceb4dd0a..1b0593b3605 100644
--- a/packages/paperclip-runner/devtools/browser/src/components/ui/badge.tsx
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/badge.tsx
@@ -6,7 +6,7 @@ export function Badge({
   tone = "neutral",
   ...props
 }: React.ComponentProps<"span"> & {
-  tone?: "neutral" | "success" | "warning" | "danger";
+  tone?: "neutral" | "accent" | "success" | "warning" | "danger";
 }) {
   return (
     <span
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/banner.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/banner.tsx
new file mode 100644
index 00000000000..7c43e7e803c
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/banner.tsx
@@ -0,0 +1,40 @@
+import * as React from "react";
+
+/**
+ * Source-adapted shadcn Alert. Failures stay in the transcript record, so this
+ * is only used for standing conditions (reconnect, pending request, replay).
+ */
+export function Banner({
+  className = "",
+  tone = "neutral",
+  assertive = false,
+  ...props
+}: React.ComponentProps<"div"> & {
+  tone?: "neutral" | "accent" | "warning" | "danger";
+  assertive?: boolean;
+}) {
+  return (
+    <div
+      data-slot="banner"
+      data-tone={tone}
+      role={assertive ? "alert" : "status"}
+      aria-live={assertive ? "assertive" : "polite"}
+      className={`ui-banner ${className}`.trim()}
+      {...props}
+    />
+  );
+}
+
+export function BannerText({ className = "", ...props }: React.ComponentProps<"p">) {
+  return <p data-slot="banner-text" className={`ui-banner-text ${className}`.trim()} {...props} />;
+}
+
+export function BannerActions({ className = "", ...props }: React.ComponentProps<"div">) {
+  return (
+    <div
+      data-slot="banner-actions"
+      className={`ui-banner-actions ${className}`.trim()}
+      {...props}
+    />
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/composer.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/composer.tsx
new file mode 100644
index 00000000000..a9360b58b35
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/composer.tsx
@@ -0,0 +1,133 @@
+import * as React from "react";
+
+import type { ComposerState } from "../../live/transcript-model";
+import { Button } from "./button";
+
+const MAX_ROWS = 8;
+
+const PRIMARY_LABEL: Record<ComposerState, string> = {
+  idle: "Send",
+  submitting: "Sending…",
+  "active-turn": "Steer",
+  interrupting: "Stopping…",
+  disconnected: "Send",
+  terminal: "Send",
+};
+
+const HINT: Record<ComposerState, string> = {
+  idle: "Enter sends. Shift and Enter add a line.",
+  submitting: "Waiting for the turn to be accepted.",
+  "active-turn": "Enter steers the running turn. Escape moves focus to Stop.",
+  interrupting: "Stopping the turn. Waiting for the terminal turn event.",
+  disconnected: "The composer is disabled until the connection is restored.",
+  terminal: "This session is closed. Replay it or reset the demo state.",
+};
+
+/**
+ * Source-adapted AI Elements `PromptInput`. The textarea plus action-row
+ * anatomy and submit-on-Enter are kept; the Send/Steer/Stop tri-state comes
+ * from the runner protocol, and there are no attachments or model pickers.
+ */
+export function Composer({
+  state,
+  value,
+  onValueChange,
+  onSubmit,
+  onStop,
+  disabledReason = null,
+  label = "Message",
+}: {
+  state: ComposerState;
+  value: string;
+  onValueChange: (value: string) => void;
+  onSubmit: () => void;
+  onStop: () => void;
+  disabledReason?: string | null;
+  label?: string;
+}) {
+  const textareaRef = React.useRef<HTMLTextAreaElement>(null);
+  const stopRef = React.useRef<HTMLButtonElement>(null);
+  const inputId = React.useId();
+  const hintId = React.useId();
+  const reasonId = React.useId();
+
+  const locked =
+    disabledReason !== null ||
+    state === "submitting" ||
+    state === "interrupting" ||
+    state === "disconnected" ||
+    state === "terminal";
+  const canStop = state === "active-turn" || state === "submitting";
+
+  React.useEffect(() => {
+    const textarea = textareaRef.current;
+    if (textarea === null) return;
+    textarea.rows = 1;
+    const rows = Math.min(value.split("\n").length, MAX_ROWS);
+    textarea.rows = Math.max(rows, 2);
+  }, [value]);
+
+  function onKeyDown(event: React.KeyboardEvent<HTMLTextAreaElement>) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      if (!locked && value.trim().length > 0) onSubmit();
+      return;
+    }
+    // Escape never interrupts directly; it only moves focus to Stop.
+    if (event.key === "Escape" && canStop) {
+      event.preventDefault();
+      stopRef.current?.focus();
+    }
+  }
+
+  return (
+    <div data-slot="composer" data-state={state} className="ui-composer">
+      <label className="ui-visually-hidden" htmlFor={inputId}>
+        {label}
+      </label>
+      <textarea
+        ref={textareaRef}
+        id={inputId}
+        className="ui-composer-input"
+        data-testid="composer-input"
+        rows={2}
+        value={value}
+        disabled={locked}
+        aria-describedby={disabledReason === null ? hintId : `${hintId} ${reasonId}`}
+        placeholder={state === "active-turn" ? "Steer the running turn…" : "Send a message…"}
+        onChange={(event) => onValueChange(event.target.value)}
+        onKeyDown={onKeyDown}
+      />
+      <div className="ui-composer-actions">
+        <p id={hintId} className="ui-composer-hint">
+          {HINT[state]}
+        </p>
+        {disabledReason === null ? null : (
+          <p id={reasonId} className="ui-composer-reason" data-testid="composer-reason">
+            {disabledReason}
+          </p>
+        )}
+        <div className="ui-composer-buttons">
+          <Button
+            ref={stopRef}
+            type="button"
+            className="ui-button--danger"
+            data-testid="composer-stop"
+            disabled={!canStop}
+            onClick={onStop}
+          >
+            Stop
+          </Button>
+          <Button
+            type="button"
+            data-testid="composer-send"
+            disabled={locked || value.trim().length === 0}
+            onClick={onSubmit}
+          >
+            {PRIMARY_LABEL[state]}
+          </Button>
+        </div>
+      </div>
+    </div>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/conversation.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/conversation.tsx
new file mode 100644
index 00000000000..8f3a4782c3c
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/conversation.tsx
@@ -0,0 +1,95 @@
+import * as React from "react";
+
+/**
+ * Source-adapted AI Elements `Conversation`. The stick-to-bottom semantics and
+ * the "jump to latest" affordance are kept; the `use-stick-to-bottom`
+ * dependency is replaced by a scroll listener, and the unseen count comes from
+ * the reducer timeline length rather than a message store.
+ */
+export function Conversation({
+  itemCount,
+  label,
+  children,
+}: {
+  itemCount: number;
+  label: string;
+  children: React.ReactNode;
+}) {
+  const viewportRef = React.useRef<HTMLDivElement>(null);
+  const [following, setFollowing] = React.useState(true);
+  const [seenCount, setSeenCount] = React.useState(itemCount);
+
+  const scrollToBottom = React.useCallback((smooth: boolean) => {
+    const viewport = viewportRef.current;
+    if (viewport === null) return;
+    const reduceMotion = window.matchMedia?.("(prefers-reduced-motion: reduce)").matches === true;
+    viewport.scrollTo({
+      top: viewport.scrollHeight,
+      behavior: smooth && !reduceMotion ? "smooth" : "auto",
+    });
+  }, []);
+
+  React.useEffect(() => {
+    if (!following) return;
+    scrollToBottom(true);
+    setSeenCount(itemCount);
+  }, [itemCount, following, scrollToBottom]);
+
+  function onScroll() {
+    const viewport = viewportRef.current;
+    if (viewport === null) return;
+    const atBottom =
+      viewport.scrollHeight - viewport.scrollTop - viewport.clientHeight < 24;
+    setFollowing(atBottom);
+    if (atBottom) setSeenCount(itemCount);
+  }
+
+  function reengage() {
+    setFollowing(true);
+    setSeenCount(itemCount);
+    scrollToBottom(true);
+  }
+
+  function onKeyDown(event: React.KeyboardEvent<HTMLDivElement>) {
+    if (event.key === "End") {
+      event.preventDefault();
+      reengage();
+    } else if (event.key === "Home") {
+      event.preventDefault();
+      setFollowing(false);
+      viewportRef.current?.scrollTo({ top: 0 });
+    }
+  }
+
+  const unseen = Math.max(itemCount - seenCount, 0);
+
+  return (
+    <div data-slot="conversation" className="ui-conversation">
+      <div
+        ref={viewportRef}
+        role="log"
+        aria-label={label}
+        aria-live="polite"
+        aria-relevant="additions"
+        tabIndex={0}
+        className="ui-conversation-viewport"
+        data-following={following}
+        data-testid="live-transcript"
+        onScroll={onScroll}
+        onKeyDown={onKeyDown}
+      >
+        {children}
+      </div>
+      {following ? null : (
+        <button
+          type="button"
+          className="ui-jump-to-latest"
+          data-testid="jump-to-latest"
+          onClick={reengage}
+        >
+          Jump to latest{unseen > 0 ? ` (${unseen} new)` : ""}
+        </button>
+      )}
+    </div>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/dialog.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/dialog.tsx
new file mode 100644
index 00000000000..120a4d86845
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/dialog.tsx
@@ -0,0 +1,61 @@
+import * as React from "react";
+
+/**
+ * Source-adapted shadcn Dialog rebuilt on the native `<dialog>` element.
+ * `showModal()` supplies the focus trap, the backdrop, and Escape-to-close,
+ * so no radix runtime and no focus-trap dependency are needed.
+ */
+export function Dialog({
+  open,
+  onClose,
+  title,
+  description,
+  children,
+  footer,
+}: {
+  open: boolean;
+  onClose: () => void;
+  title: string;
+  description?: string;
+  children?: React.ReactNode;
+  footer: React.ReactNode;
+}) {
+  const ref = React.useRef<HTMLDialogElement>(null);
+  const titleId = React.useId();
+  const descriptionId = React.useId();
+
+  React.useEffect(() => {
+    const element = ref.current;
+    if (element === null) return;
+    if (open && !element.open) element.showModal();
+    if (!open && element.open) element.close();
+  }, [open]);
+
+  return (
+    <dialog
+      ref={ref}
+      data-slot="dialog"
+      className="ui-dialog"
+      aria-labelledby={titleId}
+      {...(description === undefined ? {} : { "aria-describedby": descriptionId })}
+      onCancel={(event) => {
+        event.preventDefault();
+        onClose();
+      }}
+      onClose={onClose}
+    >
+      <form method="dialog" className="ui-dialog-body">
+        <h2 id={titleId} className="ui-dialog-title">
+          {title}
+        </h2>
+        {description === undefined ? null : (
+          <p id={descriptionId} className="ui-dialog-description">
+            {description}
+          </p>
+        )}
+        {children}
+        <div className="ui-dialog-footer">{footer}</div>
+      </form>
+    </dialog>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/menu.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/menu.tsx
new file mode 100644
index 00000000000..b5645de0b87
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/menu.tsx
@@ -0,0 +1,138 @@
+import * as React from "react";
+
+export interface MenuItemDefinition {
+  id: string;
+  label: string;
+  enabled: boolean;
+  /** Exact upstream diagnostic; exposed through aria-describedby, not hover. */
+  reason: string | null;
+}
+
+/**
+ * Source-adapted shadcn DropdownMenu narrowed to the menu-button pattern the
+ * goal surface needs. Five fixed verbs, no search, no radix portal.
+ */
+export function Menu({
+  label,
+  items,
+  onSelect,
+  disabled = false,
+  disabledReason = null,
+  triggerId,
+}: {
+  label: string;
+  items: readonly MenuItemDefinition[];
+  onSelect: (id: string) => void;
+  disabled?: boolean;
+  disabledReason?: string | null;
+  triggerId: string;
+}) {
+  const [open, setOpen] = React.useState(false);
+  const containerRef = React.useRef<HTMLDivElement>(null);
+  const triggerRef = React.useRef<HTMLButtonElement>(null);
+  const descriptionId = `${triggerId}-reason`;
+
+  React.useEffect(() => {
+    if (!open) return undefined;
+    function onPointerDown(event: MouseEvent) {
+      if (!containerRef.current?.contains(event.target as Node)) setOpen(false);
+    }
+    document.addEventListener("mousedown", onPointerDown);
+    return () => document.removeEventListener("mousedown", onPointerDown);
+  }, [open]);
+
+  function focusItem(index: number) {
+    const buttons = containerRef.current?.querySelectorAll<HTMLButtonElement>(
+      '[role="menuitem"]',
+    );
+    if (buttons === undefined || buttons.length === 0) return;
+    buttons[(index + buttons.length) % buttons.length]?.focus();
+  }
+
+  function onTriggerKeyDown(event: React.KeyboardEvent<HTMLButtonElement>) {
+    if (["ArrowDown", "Enter", " "].includes(event.key)) {
+      event.preventDefault();
+      setOpen(true);
+      window.setTimeout(() => focusItem(0), 0);
+    }
+  }
+
+  function onMenuKeyDown(event: React.KeyboardEvent<HTMLDivElement>) {
+    const buttons = [
+      ...(containerRef.current?.querySelectorAll<HTMLButtonElement>('[role="menuitem"]') ?? []),
+    ];
+    const index = buttons.indexOf(document.activeElement as HTMLButtonElement);
+    if (event.key === "Escape") {
+      event.preventDefault();
+      setOpen(false);
+      triggerRef.current?.focus();
+    } else if (event.key === "ArrowDown") {
+      event.preventDefault();
+      focusItem(index + 1);
+    } else if (event.key === "ArrowUp") {
+      event.preventDefault();
+      focusItem(index - 1);
+    }
+  }
+
+  return (
+    <div ref={containerRef} data-slot="menu" className="ui-menu">
+      <button
+        ref={triggerRef}
+        id={triggerId}
+        type="button"
+        className="ui-button ui-button--quiet"
+        aria-haspopup="menu"
+        aria-expanded={open}
+        aria-disabled={disabled}
+        {...(disabled && disabledReason !== null ? { "aria-describedby": descriptionId } : {})}
+        onClick={() => {
+          if (disabled) return;
+          setOpen((current) => !current);
+        }}
+        onKeyDown={(event) => {
+          if (disabled) return;
+          onTriggerKeyDown(event);
+        }}
+      >
+        {label}
+        <span aria-hidden="true"> ▾</span>
+      </button>
+      {disabled && disabledReason !== null ? (
+        <p id={descriptionId} className="ui-menu-reason">
+          {disabledReason}
+        </p>
+      ) : null}
+      {open && !disabled ? (
+        <div role="menu" aria-labelledby={triggerId} className="ui-menu-list" onKeyDown={onMenuKeyDown}>
+          {items.map((item) => {
+            const itemReasonId = `${triggerId}-${item.id}-reason`;
+            return (
+              <React.Fragment key={item.id}>
+                <button
+                  type="button"
+                  role="menuitem"
+                  className="ui-menu-item"
+                  aria-disabled={!item.enabled}
+                  {...(item.reason === null ? {} : { "aria-describedby": itemReasonId })}
+                  onClick={() => {
+                    if (!item.enabled) return;
+                    setOpen(false);
+                    onSelect(item.id);
+                  }}
+                >
+                  {item.label}
+                </button>
+                {item.reason === null ? null : (
+                  <p id={itemReasonId} className="ui-menu-reason">
+                    {item.reason}
+                  </p>
+                )}
+              </React.Fragment>
+            );
+          })}
+        </div>
+      ) : null}
+    </div>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/message.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/message.tsx
new file mode 100644
index 00000000000..4ad02971d81
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/message.tsx
@@ -0,0 +1,60 @@
+import * as React from "react";
+
+import type { TranscriptRole } from "../../live/transcript-model";
+
+const ROLE_LABELS: Record<TranscriptRole, string> = {
+  user: "You",
+  assistant: "Assistant",
+  reasoning: "Reasoning",
+  tool: "Tool",
+  steering: "Steering acknowledged",
+  interrupt: "Interrupt acknowledged",
+  goal: "Goal",
+  lineage: "Child thread",
+  system: "System",
+};
+
+/**
+ * Source-adapted AI Elements `Message`. The role-based alignment and the
+ * compact avatar-less variant are kept; the props are reducer item snapshots
+ * rather than AI SDK message parts.
+ */
+export function Message({
+  role,
+  children,
+  interrupted = false,
+  failed = false,
+  streaming = false,
+}: {
+  role: TranscriptRole;
+  children: React.ReactNode;
+  interrupted?: boolean;
+  failed?: boolean;
+  streaming?: boolean;
+}) {
+  return (
+    <article
+      data-slot="message"
+      data-role={role}
+      data-state={failed ? "failed" : interrupted ? "interrupted" : streaming ? "streaming" : "settled"}
+      className="ui-message"
+    >
+      <p className="ui-message-role">{ROLE_LABELS[role]}</p>
+      <div className="ui-message-body">{children}</div>
+      {interrupted ? (
+        <p className="ui-message-divider" data-testid="interrupted-divider">
+          Interrupted — the text above is everything that arrived.
+        </p>
+      ) : null}
+    </article>
+  );
+}
+
+export function MessageText({ text, streaming }: { text: string; streaming: boolean }) {
+  return (
+    <p className="ui-message-text">
+      {text}
+      {streaming ? <span className="ui-stream-cursor" aria-hidden="true" /> : null}
+    </p>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/reasoning-item.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/reasoning-item.tsx
new file mode 100644
index 00000000000..6b103c9d974
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/reasoning-item.tsx
@@ -0,0 +1,33 @@
+import * as React from "react";
+
+/**
+ * Source-adapted AI Elements `Reasoning`. The collapsible block that opens
+ * while streaming is kept; it is rebuilt on native `<details>` and renders
+ * exact streamed text with no markdown pipeline.
+ */
+export function ReasoningItem({
+  text,
+  streaming,
+  interrupted,
+}: {
+  text: string;
+  streaming: boolean;
+  interrupted: boolean;
+}) {
+  const summary = text.split("\n")[0]?.slice(0, 96) ?? "Reasoning";
+  return (
+    <details
+      data-slot="reasoning"
+      className="ui-disclosure"
+      open={streaming}
+      data-testid="reasoning-item"
+    >
+      <summary className="ui-disclosure-summary">
+        <span>{streaming ? "Reasoning…" : "Reasoning"}</span>
+        <span className="ui-disclosure-preview">{summary}</span>
+      </summary>
+      <pre className="ui-payload">{text}</pre>
+      {interrupted ? <p className="ui-message-divider">Interrupted while reasoning.</p> : null}
+    </details>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/tabs.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/tabs.tsx
new file mode 100644
index 00000000000..b017b522b1a
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/tabs.tsx
@@ -0,0 +1,94 @@
+import * as React from "react";
+
+export interface TabDefinition {
+  id: string;
+  label: string;
+}
+
+/**
+ * Source-adapted shadcn Tabs rebuilt on the WAI-ARIA tabs pattern with a
+ * roving tabindex. No radix runtime: arrow keys move, Home/End jump, Tab
+ * enters the panel.
+ */
+export function Tabs({
+  tabs,
+  value,
+  onValueChange,
+  label,
+  className = "",
+  children,
+}: {
+  tabs: readonly TabDefinition[];
+  value: string;
+  onValueChange: (value: string) => void;
+  label: string;
+  className?: string;
+  children: React.ReactNode;
+}) {
+  const listRef = React.useRef<HTMLDivElement>(null);
+
+  function moveFocus(nextIndex: number) {
+    const next = tabs[(nextIndex + tabs.length) % tabs.length];
+    if (next === undefined) return;
+    onValueChange(next.id);
+    const button = listRef.current?.querySelector<HTMLButtonElement>(
+      `[data-tab-id="${next.id}"]`,
+    );
+    button?.focus();
+  }
+
+  function onKeyDown(event: React.KeyboardEvent<HTMLDivElement>) {
+    const index = tabs.findIndex((tab) => tab.id === value);
+    if (event.key === "ArrowRight") {
+      event.preventDefault();
+      moveFocus(index + 1);
+    } else if (event.key === "ArrowLeft") {
+      event.preventDefault();
+      moveFocus(index - 1);
+    } else if (event.key === "Home") {
+      event.preventDefault();
+      moveFocus(0);
+    } else if (event.key === "End") {
+      event.preventDefault();
+      moveFocus(tabs.length - 1);
+    }
+  }
+
+  return (
+    <div data-slot="tabs" className={`ui-tabs ${className}`.trim()}>
+      <div
+        ref={listRef}
+        role="tablist"
+        aria-label={label}
+        className="ui-tablist"
+        onKeyDown={onKeyDown}
+      >
+        {tabs.map((tab) => (
+          <button
+            key={tab.id}
+            type="button"
+            role="tab"
+            data-tab-id={tab.id}
+            id={`tab-${tab.id}`}
+            aria-selected={tab.id === value}
+            aria-controls={`tabpanel-${tab.id}`}
+            tabIndex={tab.id === value ? 0 : -1}
+            className="ui-tab"
+            onClick={() => onValueChange(tab.id)}
+          >
+            {tab.label}
+          </button>
+        ))}
+      </div>
+      <div
+        role="tabpanel"
+        id={`tabpanel-${value}`}
+        aria-labelledby={`tab-${value}`}
+        tabIndex={0}
+        className="ui-tabpanel"
+      >
+        {children}
+      </div>
+    </div>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/tool-item.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/tool-item.tsx
new file mode 100644
index 00000000000..8de41978971
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/tool-item.tsx
@@ -0,0 +1,80 @@
+import * as React from "react";
+
+import { Badge } from "./badge";
+
+export type ToolStatus = "running" | "completed" | "failed" | "interrupted";
+
+const STATUS_TONE: Record<ToolStatus, "neutral" | "accent" | "success" | "danger" | "warning"> = {
+  running: "accent",
+  completed: "success",
+  failed: "danger",
+  interrupted: "warning",
+};
+
+const STATUS_LABEL: Record<ToolStatus, string> = {
+  running: "Running",
+  completed: "Completed",
+  failed: "Failed",
+  interrupted: "Interrupted",
+};
+
+/**
+ * Source-adapted AI Elements `Tool`. The header/input/output anatomy is kept;
+ * the status enum comes from canonical item events instead of an AI SDK tool
+ * part, and payloads stay in `<pre>` so nothing is prettified away.
+ */
+export function ToolItem({
+  name,
+  status,
+  input,
+  output,
+  failure,
+}: {
+  name: string;
+  status: ToolStatus;
+  input: string | null;
+  output: string | null;
+  failure: string | null;
+}) {
+  return (
+    <details
+      data-slot="tool"
+      className="ui-disclosure"
+      open={status === "running" || status === "failed"}
+      data-testid="tool-item"
+      data-status={status}
+    >
+      <summary className="ui-disclosure-summary">
+        <span>
+          Tool <code>{name}</code>
+        </span>
+        <Badge tone={STATUS_TONE[status]}>{STATUS_LABEL[status]}</Badge>
+      </summary>
+      {input === null ? null : (
+        <>
+          <p className="ui-disclosure-label">Input</p>
+          <pre className="ui-payload">{input}</pre>
+        </>
+      )}
+      {output === null ? null : (
+        <>
+          <p className="ui-disclosure-label">Output</p>
+          <pre className="ui-payload">{output}</pre>
+        </>
+      )}
+      {failure === null ? null : (
+        <>
+          <p className="ui-disclosure-label">Diagnostic</p>
+          <pre className="ui-payload ui-payload--danger" data-testid="tool-failure">
+            {failure}
+          </pre>
+        </>
+      )}
+      {status === "interrupted" ? (
+        <p className="ui-message-divider" data-testid="interrupted-divider">
+          Interrupted — the last known status is shown and no result was fabricated.
+        </p>
+      ) : null}
+    </details>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/components/ui/tooltip.tsx b/packages/paperclip-runner/devtools/browser/src/components/ui/tooltip.tsx
new file mode 100644
index 00000000000..6f367d8adae
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/components/ui/tooltip.tsx
@@ -0,0 +1,25 @@
+import * as React from "react";
+
+/**
+ * Source-adapted shadcn Tooltip. The content is always mirrored into a
+ * described-by element, so a tooltip is never the only channel for a
+ * diagnostic (interaction map A5).
+ */
+export function Tooltip({
+  content,
+  id,
+  children,
+}: {
+  content: string;
+  id: string;
+  children: React.ReactElement<{ "aria-describedby"?: string }>;
+}) {
+  return (
+    <span data-slot="tooltip" className="ui-tooltip">
+      {React.cloneElement(children, { "aria-describedby": id })}
+      <span role="tooltip" id={id} className="ui-tooltip-content">
+        {content}
+      </span>
+    </span>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/Inspector.tsx b/packages/paperclip-runner/devtools/browser/src/live/Inspector.tsx
new file mode 100644
index 00000000000..90dbc8abf5c
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/Inspector.tsx
@@ -0,0 +1,229 @@
+import * as React from "react";
+
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import type { SessionSnapshot } from "../../../../src/reducer/session-reducer";
+import { Badge } from "../components/ui/badge";
+import { Tabs } from "../components/ui/tabs";
+import type { LiveConnectionStatus, LiveSessionState } from "./protocol";
+import { capabilityRows } from "./transcript-model";
+
+const TABS = [
+  { id: "events", label: "Events" },
+  { id: "requests", label: "Requests" },
+  { id: "capabilities", label: "Capabilities" },
+  { id: "session", label: "Session" },
+];
+
+function CopyButton({ label, value }: { label: string; value: string }) {
+  const [copied, setCopied] = React.useState(false);
+  return (
+    <button
+      type="button"
+      className="ui-copy"
+      aria-label={label}
+      onClick={() => {
+        void navigator.clipboard?.writeText(value).then(
+          () => setCopied(true),
+          () => setCopied(false),
+        );
+      }}
+    >
+      {copied ? "Copied" : "Copy"}
+    </button>
+  );
+}
+
+/**
+ * The debugging surface. Redaction is the server's job: the inspector renders
+ * redaction markers verbatim and never tries to reconstruct a redacted value.
+ */
+export function Inspector({
+  events,
+  snapshot,
+  state,
+  connection,
+  reconnectAttempt,
+  replayParity,
+  highlightItemId,
+  onHighlightItem,
+}: {
+  events: readonly PrpEvent[];
+  snapshot: SessionSnapshot | null;
+  state: LiveSessionState | null;
+  connection: LiveConnectionStatus;
+  reconnectAttempt: number;
+  replayParity: boolean | null;
+  highlightItemId: string | null;
+  onHighlightItem: (itemId: string | null) => void;
+}) {
+  const [tab, setTab] = React.useState("events");
+  const [filter, setFilter] = React.useState("");
+  const [facet, setFacet] = React.useState<string | null>(null);
+
+  const facets = React.useMemo(
+    () => [...new Set(events.map((event) => event.eventType))].sort(),
+    [events],
+  );
+
+  const visible = React.useMemo(
+    () =>
+      events.filter((event) => {
+        if (facet !== null && event.eventType !== facet) return false;
+        if (filter.length === 0) return true;
+        return JSON.stringify(event).toLowerCase().includes(filter.toLowerCase());
+      }),
+    [events, facet, filter],
+  );
+
+  return (
+    <Tabs tabs={TABS} value={tab} onValueChange={setTab} label="Protocol inspector">
+      {tab === "events" ? (
+        <div className="ui-inspector-events">
+          <label className="ui-visually-hidden" htmlFor="inspector-filter">
+            Filter events
+          </label>
+          <input
+            id="inspector-filter"
+            className="ui-inspector-filter"
+            data-testid="inspector-filter"
+            placeholder="Filter events"
+            value={filter}
+            onChange={(event) => setFilter(event.target.value)}
+          />
+          <div className="ui-facets">
+            <button
+              type="button"
+              className="ui-facet"
+              aria-pressed={facet === null}
+              onClick={() => setFacet(null)}
+            >
+              All
+            </button>
+            {facets.map((candidate) => (
+              <button
+                key={candidate}
+                type="button"
+                className="ui-facet"
+                aria-pressed={facet === candidate}
+                onClick={() => setFacet(candidate)}
+              >
+                {candidate}
+              </button>
+            ))}
+          </div>
+          <ol className="ui-inspector-list" data-testid="inspector-events">
+            {visible.map((event) => (
+              <li
+                key={event.sourceEventId}
+                data-highlighted={
+                  highlightItemId !== null && event.itemId === highlightItemId
+                }
+              >
+                <details
+                  onToggle={(toggle) =>
+                    onHighlightItem(
+                      (toggle.currentTarget as HTMLDetailsElement).open
+                        ? (event.itemId ?? null)
+                        : null,
+                    )
+                  }
+                >
+                  <summary className="ui-inspector-row">
+                    <code className="ui-seq">{event.sourceSeq}</code>
+                    <code>{event.eventType}</code>
+                    <code className="ui-inspector-id">{event.itemId ?? event.turnId ?? "—"}</code>
+                  </summary>
+                  <pre className="ui-payload">{JSON.stringify(event, null, 2)}</pre>
+                  <CopyButton
+                    label={`Copy event ${event.sourceSeq} payload`}
+                    value={JSON.stringify(event, null, 2)}
+                  />
+                </details>
+              </li>
+            ))}
+          </ol>
+        </div>
+      ) : null}
+
+      {tab === "requests" ? (
+        <dl className="ui-inspector-grid" data-testid="inspector-requests">
+          {(snapshot?.requests ?? []).map((request) => (
+            <div key={request.requestId}>
+              <dt>
+                <code>{request.requestId}</code>
+              </dt>
+              <dd>
+                {request.type} — {request.status}
+              </dd>
+            </div>
+          ))}
+          {(snapshot?.requests.length ?? 0) === 0 ? <p>No runtime requests yet.</p> : null}
+        </dl>
+      ) : null}
+
+      {tab === "capabilities" ? (
+        <ul className="ui-capability-list" data-testid="inspector-capabilities">
+          {capabilityRows(state?.capabilities ?? { resume: false, typedEvents: false, steering: false, interruption: false, structuredResult: false }).map(
+            (row) => (
+              <li key={row.name}>
+                <span>{row.name}</span>
+                <Badge tone={row.supported ? "success" : "warning"}>
+                  {row.supported ? "supported" : "unsupported"}
+                </Badge>
+                {row.diagnostic === null ? null : (
+                  <p className="ui-capability-diagnostic">{row.diagnostic}</p>
+                )}
+              </li>
+            ),
+          )}
+        </ul>
+      ) : null}
+
+      {tab === "session" ? (
+        <dl className="ui-inspector-grid" data-testid="inspector-session">
+          <div>
+            <dt>Run</dt>
+            <dd>{state?.runId ?? "—"}</dd>
+          </div>
+          <div>
+            <dt>Normalized session</dt>
+            <dd>{state?.normalizedSessionId ?? "—"}</dd>
+          </div>
+          <div>
+            <dt>Driver session</dt>
+            <dd>{state?.driverSession.driverSessionId ?? "—"}</dd>
+          </div>
+          <div>
+            <dt>Source cursors</dt>
+            <dd>{JSON.stringify(snapshot?.sourceCursors ?? {})}</dd>
+          </div>
+          <div>
+            <dt>Integrity</dt>
+            <dd>{snapshot?.integrity ?? "—"}</dd>
+          </div>
+          <div>
+            <dt>Gaps</dt>
+            <dd>{snapshot?.gaps.length ?? 0}</dd>
+          </div>
+          <div>
+            <dt>Connection</dt>
+            <dd>
+              {connection}
+              {connection === "reconnecting" ? ` (attempt ${reconnectAttempt})` : ""}
+            </dd>
+          </div>
+          <div>
+            <dt>Credentials in browser</dt>
+            <dd>{state?.credentialsExposed === true ? "yes" : "no"}</dd>
+          </div>
+          <div>
+            <dt>Live and replay reducer</dt>
+            <dd data-testid="replay-parity">
+              {replayParity === null ? "not compared yet" : replayParity ? "match" : "mismatch"}
+            </dd>
+          </div>
+        </dl>
+      ) : null}
+    </Tabs>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/LiveConsole.tsx b/packages/paperclip-runner/devtools/browser/src/live/LiveConsole.tsx
new file mode 100644
index 00000000000..0d2eb2cb230
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/LiveConsole.tsx
@@ -0,0 +1,487 @@
+import * as React from "react";
+
+import { Badge } from "../components/ui/badge";
+import { Banner, BannerActions, BannerText } from "../components/ui/banner";
+import { Button } from "../components/ui/button";
+import { Composer } from "../components/ui/composer";
+import { Conversation } from "../components/ui/conversation";
+import { Dialog } from "../components/ui/dialog";
+import { Message, MessageText } from "../components/ui/message";
+import { ReasoningItem } from "../components/ui/reasoning-item";
+import { Tabs } from "../components/ui/tabs";
+import { ToolItem, type ToolStatus } from "../components/ui/tool-item";
+import { Inspector } from "./Inspector";
+import { RequestCard } from "./RequestCard";
+import { ConnectionBlock, GoalBlock, LineageTree, ManifestList } from "./SessionPanel";
+import { useLiveConsole } from "./use-live-console";
+import type { TranscriptEntry } from "./transcript-model";
+
+const MOBILE_SEGMENTS = [
+  { id: "chat", label: "Chat" },
+  { id: "session", label: "Session" },
+  { id: "inspector", label: "Inspector" },
+];
+
+const TURN_TONE = {
+  completed: "success",
+  failed: "danger",
+  interrupted: "warning",
+  cancelled: "warning",
+} as const;
+
+const TURN_LABEL = {
+  completed: "Turn completed",
+  failed: "Turn failed",
+  interrupted: "Turn interrupted",
+  cancelled: "Cancelled before start",
+} as const;
+
+/**
+ * Renders exactly one layout. Hiding a second copy with CSS would duplicate
+ * the transcript log landmark and every control in the DOM.
+ */
+function useCompactLayout(): boolean {
+  const [compact, setCompact] = React.useState(
+    () => window.matchMedia?.("(max-width: 64rem)").matches === true,
+  );
+  React.useEffect(() => {
+    const query = window.matchMedia?.("(max-width: 64rem)");
+    if (query === undefined) return undefined;
+    const onChange = () => setCompact(query.matches);
+    onChange();
+    query.addEventListener("change", onChange);
+    return () => query.removeEventListener("change", onChange);
+  }, []);
+  return compact;
+}
+
+function toolStatus(entry: Extract<TranscriptEntry, { kind: "item" }>): ToolStatus {
+  if (entry.interrupted) return "interrupted";
+  if (entry.item.status === "failed") return "failed";
+  if (entry.item.status === "running") return "running";
+  return "completed";
+}
+
+export function LiveConsole() {
+  const console_ = useLiveConsole();
+  const [draft, setDraft] = React.useState("");
+  const [inspectorOpen, setInspectorOpen] = React.useState(false);
+  const [segment, setSegment] = React.useState("chat");
+  const [resetOpen, setResetOpen] = React.useState(false);
+  const [highlightItemId, setHighlightItemId] = React.useState<string | null>(null);
+  const [outcomes, setOutcomes] = React.useState<Record<string, string>>({});
+  const oldestPendingRef = React.useRef<HTMLLIElement>(null);
+  const compact = useCompactLayout();
+
+  React.useEffect(() => {
+    const stored = window.localStorage.getItem("paperclip-runner.phase4b.inspector");
+    if (stored === "open") setInspectorOpen(true);
+  }, []);
+
+  React.useEffect(() => {
+    window.localStorage.setItem(
+      "paperclip-runner.phase4b.inspector",
+      inspectorOpen ? "open" : "closed",
+    );
+  }, [inspectorOpen]);
+
+  const manifestId = console_.state?.manifest ?? console_.selectedManifestId;
+  const activeManifest = console_.manifests.find((entry) => entry.id === manifestId);
+  const pendingRequests = console_.transcript.filter(
+    (entry) => entry.kind === "request" && entry.status === "pending",
+  );
+  const viewingChild = console_.selectedThreadId !== null;
+  const childSteeringSupported = false;
+
+  React.useEffect(() => {
+    const snapshot = console_.snapshot;
+    if (snapshot === null || console_.state === null) return;
+    if (!["completed", "failed", "interrupted", "cancelled"].includes(snapshot.turnState)) return;
+    setOutcomes((current) => ({
+      ...current,
+      [console_.state!.manifest ?? "unknown"]:
+        snapshot.turnState === "completed" ? "passed observations" : snapshot.turnState,
+    }));
+  }, [console_.snapshot, console_.state]);
+
+  function submit() {
+    const text = draft.trim();
+    if (text.length === 0) return;
+    setDraft("");
+    if (console_.composer === "active-turn") void console_.steer(text);
+    else void console_.send(text);
+  }
+
+  function runManifest(id: string) {
+    const manifest = console_.manifests.find((entry) => entry.id === id);
+    void console_.start({ manifestId: id, message: manifest?.prompt ?? "Run the demo." });
+  }
+
+  const composerReason = viewingChild && !childSteeringSupported
+    ? "Direct steering of child threads is not supported by this app-server."
+    : null;
+
+  const transcriptBody = (
+    <Conversation itemCount={console_.transcript.length} label="Session transcript">
+      {console_.transcript.length === 0 ? (
+        <div className="ui-empty-transcript" data-testid="empty-transcript">
+          <p className="eyebrow">No session yet</p>
+          <h3>Pick a demo chat or start a blank session</h3>
+          <p>
+            The browser never holds a provider credential. Only the demo server starts a driver
+            and owns the working directory.
+          </p>
+        </div>
+      ) : null}
+      <ol className="ui-transcript" role="list">
+        {console_.transcript.map((entry) => {
+          if (entry.kind === "user") {
+            return (
+              <li key={entry.key}>
+                <Message role="user">
+                  <MessageText text={entry.text} streaming={false} />
+                </Message>
+              </li>
+            );
+          }
+          if (entry.kind === "turn") {
+            return (
+              <li key={entry.key} className="ui-turn-marker" data-testid={`turn-${entry.state}`}>
+                <Badge tone={TURN_TONE[entry.state]}>{TURN_LABEL[entry.state]}</Badge>
+                <span>{entry.detail}</span>
+              </li>
+            );
+          }
+          if (entry.kind === "diagnostic") {
+            return (
+              <li key={entry.key}>
+                <Message role="system" failed>
+                  <MessageText text={entry.message} streaming={false} />
+                </Message>
+              </li>
+            );
+          }
+          if (entry.kind === "request") {
+            const isOldestPending =
+              entry.status === "pending" && pendingRequests[0]?.key === entry.key;
+            return (
+              <RequestCard
+                key={entry.key}
+                entry={entry}
+                {...(isOldestPending ? { focusRef: oldestPendingRef } : {})}
+                onResolve={(resolution) =>
+                  void console_.resolve(entry.requestId, entry.turnId ?? "", resolution)
+                }
+              />
+            );
+          }
+          if (entry.role === "reasoning") {
+            return (
+              <li key={entry.key} data-highlighted={entry.item.itemId === highlightItemId}>
+                <ReasoningItem
+                  text={entry.item.text}
+                  streaming={entry.streaming}
+                  interrupted={entry.interrupted}
+                />
+              </li>
+            );
+          }
+          if (entry.role === "tool") {
+            return (
+              <li key={entry.key} data-highlighted={entry.item.itemId === highlightItemId}>
+                <ToolItem
+                  name={entry.toolName ?? entry.item.kind}
+                  status={toolStatus(entry)}
+                  input={entry.input}
+                  output={entry.output}
+                  failure={entry.failure}
+                />
+              </li>
+            );
+          }
+          return (
+            <li key={entry.key} data-highlighted={entry.item.itemId === highlightItemId}>
+              <Message
+                role={entry.role}
+                streaming={entry.streaming}
+                interrupted={entry.interrupted}
+                failed={entry.item.status === "failed"}
+              >
+                <MessageText text={entry.item.text} streaming={entry.streaming} />
+                {entry.failure === null ? null : (
+                  <pre className="ui-payload ui-payload--danger">{entry.failure}</pre>
+                )}
+              </Message>
+            </li>
+          );
+        })}
+        {console_.steeringChips.map((chip) => (
+          <li key={chip.id} className="ui-steering-chip" data-status={chip.status} data-testid="steering-chip">
+            <Badge
+              tone={
+                chip.status === "acknowledged"
+                  ? "success"
+                  : chip.status === "pending"
+                    ? "accent"
+                    : "danger"
+              }
+            >
+              {chip.status === "rejected"
+                ? `rejected — ${chip.detail ?? "turn already ended"}`
+                : chip.status}
+            </Badge>
+            <span>{chip.text}</span>
+            {chip.status === "rejected" || chip.status === "failed" ? (
+              <Button
+                type="button"
+                className="ui-button--quiet"
+                data-testid="steer-as-new-message"
+                onClick={() => {
+                  console_.dismissSteeringChip(chip.id);
+                  void console_.send(chip.text);
+                }}
+              >
+                Send as new message
+              </Button>
+            ) : null}
+          </li>
+        ))}
+      </ol>
+    </Conversation>
+  );
+
+  const composerBlock = (
+    <div className="ui-composer-region">
+      {console_.connection === "disconnected" || console_.connection === "reconnecting" ? (
+        <Banner tone="warning" data-testid="reconnect-banner">
+          <BannerText>
+            {console_.connection === "reconnecting"
+              ? `Connection lost — reconnecting (attempt ${console_.reconnectAttempt})…`
+              : "Connection lost — the session is still on the server."}
+          </BannerText>
+          <BannerActions>
+            <Button type="button" data-testid="retry-now" onClick={console_.retryNow}>
+              Retry now
+            </Button>
+          </BannerActions>
+        </Banner>
+      ) : null}
+      {pendingRequests.length > 0 ? (
+        <Banner tone="accent" assertive data-testid="pending-request-banner">
+          <BannerText>{pendingRequests.length} request waiting — Review</BannerText>
+          <BannerActions>
+            <Button
+              type="button"
+              onClick={() => oldestPendingRef.current?.focus()}
+              data-testid="review-request"
+            >
+              Review
+            </Button>
+          </BannerActions>
+        </Banner>
+      ) : null}
+      <Composer
+        state={console_.composer}
+        value={draft}
+        onValueChange={setDraft}
+        onSubmit={submit}
+        onStop={() => void console_.interrupt()}
+        disabledReason={composerReason}
+      />
+    </div>
+  );
+
+  const sessionRail = (
+    <>
+      <ManifestList
+        manifests={console_.manifests}
+        selectedId={console_.selectedManifestId}
+        onSelect={console_.selectManifest}
+        onRun={runManifest}
+        lastOutcome={outcomes}
+        disabled={console_.busy}
+      />
+      <ConnectionBlock
+        connection={console_.connection}
+        reconnectAttempt={console_.reconnectAttempt}
+        state={console_.state}
+        pendingRequestCount={pendingRequests.length}
+        onDrop={console_.dropConnection}
+        onRetry={console_.retryNow}
+        onScrollToOldestRequest={() => oldestPendingRef.current?.focus()}
+      />
+      <GoalBlock state={console_.state} onGoal={(operation, body) => void console_.goal(operation, body)} />
+      <LineageTree
+        lineage={console_.state?.lineage ?? []}
+        selectedThreadId={console_.selectedThreadId}
+        onSelect={console_.selectThread}
+      />
+      <section className="ui-panel">
+        <h2 className="ui-panel-title">Demo state</h2>
+        <Button type="button" className="ui-button--quiet" onClick={() => setResetOpen(true)}>
+          Reset demo state
+        </Button>
+        {activeManifest === undefined ? null : (
+          <ol className="ui-observation-list" data-testid="observation-checklist">
+            {activeManifest.expectedObservations.map((observation) => (
+              <li key={observation}>
+                <label>
+                  <input type="checkbox" /> {observation}
+                </label>
+              </li>
+            ))}
+          </ol>
+        )}
+      </section>
+    </>
+  );
+
+  const inspectorPanel = (
+    <Inspector
+      events={console_.events}
+      snapshot={console_.snapshot}
+      state={console_.state}
+      connection={console_.connection}
+      reconnectAttempt={console_.reconnectAttempt}
+      replayParity={console_.replayParity}
+      highlightItemId={highlightItemId}
+      onHighlightItem={setHighlightItemId}
+    />
+  );
+
+  return (
+    <div className="live-console" data-inspector={inspectorOpen}>
+      <div className="live-console-header">
+        <div className="live-console-title">
+          {console_.replay.active ? (
+            <Badge tone="warning" data-testid="replay-badge">
+              REPLAY
+            </Badge>
+          ) : null}
+          <h2>{activeManifest?.name ?? "Live console"}</h2>
+          {viewingChild ? (
+            <p className="ui-breadcrumb" data-testid="breadcrumb">
+              root ▸ {console_.selectedThreadId}
+            </p>
+          ) : null}
+        </div>
+        <div className="live-console-header-actions">
+          <Button
+            type="button"
+            className="ui-button--quiet"
+            data-testid="toggle-inspector"
+            aria-pressed={inspectorOpen}
+            onClick={() => setInspectorOpen((current) => !current)}
+          >
+            Protocol inspector
+          </Button>
+          <Button
+            type="button"
+            className="ui-button--quiet"
+            data-testid="toggle-replay"
+            aria-pressed={console_.replay.active}
+            disabled={console_.events.length === 0}
+            onClick={() => (console_.replay.active ? console_.replay.exit() : console_.replay.enter())}
+          >
+            {console_.replay.active ? "Exit replay" : "Replay"}
+          </Button>
+        </div>
+      </div>
+
+      <p className="ui-visually-hidden" aria-live="polite" data-testid="live-announcement">
+        {console_.announcement}
+      </p>
+      {console_.error === null ? null : (
+        <Banner tone="danger" assertive data-testid="console-error">
+          <BannerText>{console_.error}</BannerText>
+        </Banner>
+      )}
+
+      {console_.replay.active ? (
+        <div className="ui-replay-stepper" data-testid="replay-stepper">
+          <Button type="button" onClick={console_.replay.togglePlay}>
+            {console_.replay.playing ? "Pause" : "Play"}
+          </Button>
+          <Button type="button" className="ui-button--quiet" onClick={() => console_.replay.step(-1)}>
+            Step back
+          </Button>
+          <Button type="button" className="ui-button--quiet" onClick={() => console_.replay.step(1)}>
+            Step forward
+          </Button>
+          <label htmlFor="replay-position">Position</label>
+          <input
+            id="replay-position"
+            type="range"
+            min={0}
+            max={console_.replay.length}
+            value={console_.replay.position}
+            data-testid="replay-position"
+            onChange={(event) => console_.replay.setPosition(Number(event.target.value))}
+          />
+          <span>
+            {console_.replay.position} / {console_.replay.length}
+          </span>
+        </div>
+      ) : null}
+
+      {compact ? (
+        <div className="live-console-mobile">
+          <Tabs
+            tabs={MOBILE_SEGMENTS}
+            value={segment}
+            onValueChange={setSegment}
+            label="Console sections"
+          >
+            {segment === "chat" ? (
+              <>
+                {transcriptBody}
+                {composerBlock}
+              </>
+            ) : null}
+            {segment === "session" ? sessionRail : null}
+            {segment === "inspector" ? inspectorPanel : null}
+          </Tabs>
+        </div>
+      ) : (
+        <div className="live-console-desktop">
+          <aside className="live-console-rail">{sessionRail}</aside>
+          <main className="live-console-center">
+            {transcriptBody}
+            {composerBlock}
+          </main>
+          {inspectorOpen ? (
+            <aside className="live-console-inspector" aria-label="Protocol inspector">
+              {inspectorPanel}
+            </aside>
+          ) : null}
+        </div>
+      )}
+
+      <Dialog
+        open={resetOpen}
+        onClose={() => setResetOpen(false)}
+        title="Reset demo state"
+        description="Discards the current live session transcript. Recorded evidence files are not touched."
+        footer={
+          <>
+            <Button type="button" className="ui-button--quiet" onClick={() => setResetOpen(false)}>
+              Keep the session
+            </Button>
+            <Button
+              type="button"
+              className="ui-button--danger"
+              data-testid="reset-confirm"
+              onClick={() => {
+                setResetOpen(false);
+                setDraft("");
+                void console_.reset();
+              }}
+            >
+              Reset demo state
+            </Button>
+          </>
+        }
+      />
+    </div>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/RequestCard.tsx b/packages/paperclip-runner/devtools/browser/src/live/RequestCard.tsx
new file mode 100644
index 00000000000..15c07540038
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/RequestCard.tsx
@@ -0,0 +1,168 @@
+import * as React from "react";
+
+import { Badge } from "../components/ui/badge";
+import { Button } from "../components/ui/button";
+import type { TranscriptRequestEntry } from "./transcript-model";
+
+const KIND_LABELS: Record<string, string> = {
+  command_approval: "Command",
+  file_approval: "File change",
+  permission_approval: "Permission",
+  user_input: "Input",
+  elicitation: "Input",
+};
+
+const ACTION_LABELS: Record<string, string> = {
+  accept: "Approve",
+  accept_for_session: "Approve for session",
+  decline: "Reject",
+  cancel: "Cancel",
+  submit: "Submit",
+};
+
+const STATUS_TONE = {
+  pending: "accent",
+  resolved: "success",
+  expired: "warning",
+  cancelled: "warning",
+} as const;
+
+function detailBlock(details: Record<string, unknown>): string | null {
+  const command = details.command;
+  if (typeof command === "string") return command;
+  const diff = details.diff;
+  if (typeof diff === "string") {
+    const path = typeof details.path === "string" ? `${details.path}\n` : "";
+    return `${path}${diff}`;
+  }
+  return null;
+}
+
+/**
+ * One card renders all five request kinds. Actions come from the upstream
+ * request contract, so the card can never invent an affordance the runner
+ * cannot honour, and the first click locks the row until the canonical
+ * resolved event arrives.
+ */
+export function RequestCard({
+  entry,
+  onResolve,
+  focusRef,
+}: {
+  entry: TranscriptRequestEntry;
+  onResolve: (resolution: Record<string, unknown>) => void;
+  focusRef?: React.Ref<HTMLLIElement>;
+}) {
+  const [answer, setAnswer] = React.useState("");
+  const [resolving, setResolving] = React.useState(false);
+  const answerId = React.useId();
+  const kindLabel = KIND_LABELS[entry.requestKind] ?? "Request";
+  const pending = entry.status === "pending";
+  const isInput = entry.requestKind === "user_input" || entry.requestKind === "elicitation";
+  const payload = detailBlock(entry.details);
+
+  React.useEffect(() => {
+    if (!pending) setResolving(false);
+  }, [pending]);
+
+  function submit(resolution: Record<string, unknown>) {
+    setResolving(true);
+    onResolve(resolution);
+  }
+
+  if (!pending) {
+    return (
+      <li
+        ref={focusRef}
+        className="ui-request ui-request--settled"
+        data-status={entry.status}
+        data-testid="request-card"
+        tabIndex={-1}
+      >
+        <details>
+          <summary className="ui-request-summary">
+            <span>
+              {kindLabel} · {entry.status === "resolved"
+                ? `resolved — ${ACTION_LABELS[entry.resolvedAction ?? ""] ?? entry.resolvedAction ?? "unknown"}`
+                : entry.status === "expired"
+                  ? "expired before response"
+                  : "cancelled"}
+            </span>
+            <Badge tone={STATUS_TONE[entry.status]}>{entry.status}</Badge>
+          </summary>
+          <p className="ui-request-prompt">{entry.prompt}</p>
+          <p className="ui-request-meta">
+            <code>{entry.requestId}</code>
+            {entry.resolvedAt === null ? null : <time dateTime={entry.resolvedAt}>{entry.resolvedAt}</time>}
+          </p>
+          {payload === null ? null : <pre className="ui-payload">{payload}</pre>}
+        </details>
+      </li>
+    );
+  }
+
+  return (
+    <li
+      ref={focusRef}
+      className="ui-request"
+      data-status="pending"
+      data-testid="request-card"
+      tabIndex={-1}
+    >
+      <div className="ui-request-header">
+        <span className="ui-request-kind">{kindLabel}</span>
+        <Badge tone="accent">pending</Badge>
+      </div>
+      <p className="ui-request-prompt">{entry.prompt}</p>
+      <p className="ui-request-meta">
+        <code>{entry.requestId}</code>
+      </p>
+      {payload === null ? null : <pre className="ui-payload">{payload}</pre>}
+      {isInput ? (
+        <div className="ui-request-input">
+          <label htmlFor={answerId}>Answer</label>
+          <input
+            id={answerId}
+            data-testid="request-answer"
+            value={answer}
+            disabled={resolving}
+            onChange={(event) => setAnswer(event.target.value)}
+          />
+        </div>
+      ) : null}
+      <div className="ui-request-actions">
+        {entry.actions.map((action) => (
+          <Button
+            key={action}
+            type="button"
+            className={
+              action === "decline"
+                ? "ui-button--danger"
+                : action === "cancel"
+                  ? "ui-button--quiet"
+                  : action === "accept" || action === "submit"
+                    ? ""
+                    : "ui-button--secondary"
+            }
+            disabled={resolving || (action === "submit" && answer.trim().length === 0)}
+            data-testid={`request-action-${action}`}
+            onClick={() =>
+              submit(
+                action === "submit"
+                  ? { action: "submit", answers: { answer: { answers: [answer] } } }
+                  : { action },
+              )
+            }
+          >
+            {ACTION_LABELS[action] ?? action}
+          </Button>
+        ))}
+      </div>
+      {resolving ? (
+        <p className="ui-request-resolving" role="status">
+          Resolving — waiting for the canonical resolved event.
+        </p>
+      ) : null}
+    </li>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/SessionPanel.tsx b/packages/paperclip-runner/devtools/browser/src/live/SessionPanel.tsx
new file mode 100644
index 00000000000..f5939551b95
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/SessionPanel.tsx
@@ -0,0 +1,346 @@
+import * as React from "react";
+
+import { Badge } from "../components/ui/badge";
+import { Button } from "../components/ui/button";
+import { Dialog } from "../components/ui/dialog";
+import { Menu } from "../components/ui/menu";
+import type {
+  GoalOperation,
+  LiveConnectionStatus,
+  LiveLineageEntry,
+  LiveManifestSummary,
+  LiveSessionState,
+} from "./protocol";
+import { GOAL_OPERATION_LABELS } from "./protocol";
+import { goalAvailability } from "./transcript-model";
+
+const CONNECTION_TONE = {
+  idle: "neutral",
+  connecting: "accent",
+  connected: "success",
+  reconnecting: "warning",
+  disconnected: "danger",
+  terminal: "neutral",
+} as const;
+
+const THREAD_STATE_LABEL: Record<string, string> = {
+  starting: "Starting",
+  running: "Running",
+  completed: "Completed",
+  failed: "Failed",
+};
+
+export function ManifestList({
+  manifests,
+  selectedId,
+  onSelect,
+  onRun,
+  lastOutcome,
+  disabled,
+}: {
+  manifests: readonly LiveManifestSummary[];
+  selectedId: string;
+  onSelect: (id: string) => void;
+  onRun: (id: string) => void;
+  lastOutcome: Record<string, string>;
+  disabled: boolean;
+}) {
+  return (
+    <section className="ui-panel" aria-labelledby="manifest-heading">
+      <h2 id="manifest-heading" className="ui-panel-title">
+        Demo chats
+      </h2>
+      <ul className="ui-manifest-list" data-testid="manifest-list">
+        {manifests.map((manifest) => (
+          <li key={manifest.id} data-selected={manifest.id === selectedId}>
+            <button
+              type="button"
+              className="ui-manifest-row"
+              aria-pressed={manifest.id === selectedId}
+              data-testid={`manifest-${manifest.id}`}
+              onClick={() => onSelect(manifest.id)}
+            >
+              <span className="ui-manifest-name">{manifest.name}</span>
+              <span className="ui-manifest-purpose">{manifest.purpose}</span>
+              <span className="ui-manifest-badges">
+                {manifest.scenarios.map((scenario) => (
+                  <Badge key={scenario}>{scenario}</Badge>
+                ))}
+                <Badge tone={lastOutcome[manifest.id] === undefined ? "neutral" : "success"}>
+                  {lastOutcome[manifest.id] ?? "not run"}
+                </Badge>
+              </span>
+            </button>
+            {manifest.id === selectedId ? (
+              <>
+                <details className="ui-disclosure" data-testid="expected-observations">
+                  <summary className="ui-disclosure-summary">Expected observations</summary>
+                  <ol className="ui-observation-list">
+                    {manifest.expectedObservations.map((observation) => (
+                      <li key={observation}>{observation}</li>
+                    ))}
+                  </ol>
+                </details>
+                <Button
+                  type="button"
+                  disabled={disabled}
+                  data-testid="run-manifest"
+                  onClick={() => onRun(manifest.id)}
+                >
+                  Run this demo chat
+                </Button>
+              </>
+            ) : null}
+          </li>
+        ))}
+      </ul>
+    </section>
+  );
+}
+
+export function ConnectionBlock({
+  connection,
+  reconnectAttempt,
+  state,
+  pendingRequestCount,
+  onDrop,
+  onRetry,
+  onScrollToOldestRequest,
+}: {
+  connection: LiveConnectionStatus;
+  reconnectAttempt: number;
+  state: LiveSessionState | null;
+  pendingRequestCount: number;
+  onDrop: () => void;
+  onRetry: () => void;
+  onScrollToOldestRequest: () => void;
+}) {
+  return (
+    <section className="ui-panel" aria-labelledby="connection-heading">
+      <h2 id="connection-heading" className="ui-panel-title">
+        Session
+      </h2>
+      <div className="ui-status-row">
+        <span>Connection</span>
+        <Badge tone={CONNECTION_TONE[connection]} data-testid="connection-status">
+          {connection === "reconnecting" ? `reconnecting (attempt ${reconnectAttempt})` : connection}
+        </Badge>
+      </div>
+      <div className="ui-status-row">
+        <span>Pending requests</span>
+        <button
+          type="button"
+          className="ui-chip"
+          data-testid="pending-request-chip"
+          disabled={pendingRequestCount === 0}
+          onClick={onScrollToOldestRequest}
+        >
+          {pendingRequestCount}
+        </button>
+      </div>
+      <div className="ui-status-row">
+        <span>Credentials in browser</span>
+        <Badge tone={state?.credentialsExposed === true ? "danger" : "success"}>
+          {state?.credentialsExposed === true ? "exposed" : "none"}
+        </Badge>
+      </div>
+      <dl className="ui-identity-grid">
+        <div>
+          <dt>Run</dt>
+          <dd>{state?.runId ?? "—"}</dd>
+        </div>
+        <div>
+          <dt>Session</dt>
+          <dd>{state?.normalizedSessionId ?? "—"}</dd>
+        </div>
+      </dl>
+      <div className="ui-panel-actions">
+        <Button
+          type="button"
+          className="ui-button--quiet"
+          data-testid="drop-connection"
+          disabled={state === null || connection === "disconnected"}
+          onClick={onDrop}
+        >
+          Simulate connection loss
+        </Button>
+        <Button
+          type="button"
+          className="ui-button--quiet"
+          data-testid="resume-session"
+          disabled={state === null}
+          onClick={onRetry}
+        >
+          Resume this session
+        </Button>
+      </div>
+    </section>
+  );
+}
+
+export function GoalBlock({
+  state,
+  onGoal,
+}: {
+  state: LiveSessionState | null;
+  onGoal: (operation: GoalOperation, body?: Record<string, unknown>) => void;
+}) {
+  const [dialogOpen, setDialogOpen] = React.useState(false);
+  const [objective, setObjective] = React.useState("");
+  const objectiveId = React.useId();
+  const capabilities = state?.capabilities ?? {
+    resume: false,
+    typedEvents: false,
+    steering: false,
+    interruption: false,
+    structuredResult: false,
+  };
+  const availability = goalAvailability(capabilities, state?.goal?.status ?? null);
+  const unsupported = capabilities.goals !== true;
+  const unsupportedReason = availability[0]?.reason ?? null;
+
+  return (
+    <section className="ui-panel" aria-labelledby="goal-heading">
+      <div className="ui-panel-header">
+        <h2 id="goal-heading" className="ui-panel-title">
+          Goal
+        </h2>
+        <Menu
+          triggerId="goal-menu"
+          label="Goal"
+          disabled={unsupported || state === null}
+          disabledReason={
+            state === null ? "Start a session before changing the goal." : unsupportedReason
+          }
+          items={availability.map((entry) => ({
+            id: entry.operation,
+            label: GOAL_OPERATION_LABELS[entry.operation],
+            enabled: entry.enabled,
+            reason: entry.reason,
+          }))}
+          onSelect={(id) => {
+            if (id === "set") setDialogOpen(true);
+            else onGoal(id as GoalOperation);
+          }}
+        />
+      </div>
+      {state === null ? null : state.goal === null ? (
+        <p className="ui-goal-empty" data-testid="goal-banner">
+          {unsupported ? "Goal operations are unsupported here." : "No goal set"}
+        </p>
+      ) : (
+        <div className="ui-goal-banner" data-testid="goal-banner">
+          <p className="ui-goal-objective">{state.goal.objective}</p>
+          <div className="ui-status-row">
+            <Badge tone={state.goal.status === "active" ? "success" : "warning"}>
+              {state.goal.status}
+            </Badge>
+            <time dateTime={new Date(state.goal.updatedAt).toISOString()}>
+              {new Date(state.goal.updatedAt).toISOString().slice(11, 19)}
+            </time>
+          </div>
+        </div>
+      )}
+      <Dialog
+        open={dialogOpen}
+        onClose={() => setDialogOpen(false)}
+        title="Set goal"
+        description="The goal is durable on the thread and every change is also a transcript event."
+        footer={
+          <>
+            <Button type="button" className="ui-button--quiet" onClick={() => setDialogOpen(false)}>
+              Cancel
+            </Button>
+            <Button
+              type="button"
+              data-testid="goal-set-confirm"
+              disabled={objective.trim().length === 0}
+              onClick={() => {
+                onGoal("set", { objective: objective.trim() });
+                setObjective("");
+                setDialogOpen(false);
+              }}
+            >
+              Set goal
+            </Button>
+          </>
+        }
+      >
+        <label htmlFor={objectiveId}>Objective</label>
+        <textarea
+          id={objectiveId}
+          className="ui-dialog-input"
+          data-testid="goal-objective"
+          rows={3}
+          value={objective}
+          onChange={(event) => setObjective(event.target.value)}
+        />
+      </Dialog>
+    </section>
+  );
+}
+
+export function LineageTree({
+  lineage,
+  selectedThreadId,
+  onSelect,
+}: {
+  lineage: readonly LiveLineageEntry[];
+  selectedThreadId: string | null;
+  onSelect: (threadId: string | null) => void;
+}) {
+  const rootId = lineage.find((entry) => entry.depth === 0)?.threadId ?? null;
+  const current = selectedThreadId ?? rootId;
+  const treeRef = React.useRef<HTMLUListElement>(null);
+
+  function onKeyDown(event: React.KeyboardEvent<HTMLUListElement>) {
+    const rows = [
+      ...(treeRef.current?.querySelectorAll<HTMLButtonElement>('[role="treeitem"]') ?? []),
+    ];
+    const index = rows.indexOf(document.activeElement as HTMLButtonElement);
+    if (event.key === "ArrowDown") {
+      event.preventDefault();
+      rows[Math.min(index + 1, rows.length - 1)]?.focus();
+    } else if (event.key === "ArrowUp") {
+      event.preventDefault();
+      rows[Math.max(index - 1, 0)]?.focus();
+    }
+  }
+
+  return (
+    <section className="ui-panel" aria-labelledby="lineage-heading">
+      <h2 id="lineage-heading" className="ui-panel-title">
+        Threads
+      </h2>
+      <ul
+        ref={treeRef}
+        role="tree"
+        aria-label="Session thread lineage"
+        className="ui-tree"
+        data-testid="lineage-tree"
+        onKeyDown={onKeyDown}
+      >
+        {lineage.length === 0 ? <li className="ui-tree-empty">No threads yet.</li> : null}
+        {lineage.map((entry) => (
+          <li key={entry.threadId} role="none" data-depth={entry.depth}>
+            <button
+              type="button"
+              role="treeitem"
+              aria-current={entry.threadId === current}
+              aria-selected={entry.threadId === current}
+              className="ui-tree-row"
+              data-testid={`thread-${entry.threadId}`}
+              onClick={() => onSelect(entry.depth === 0 ? null : entry.threadId)}
+            >
+              <span className="ui-state-dot" data-state={entry.status} aria-hidden="true" />
+              <code>{entry.nickname ?? entry.threadId}</code>
+              <span className="ui-tree-state">
+                {THREAD_STATE_LABEL[entry.status] ?? entry.status}
+              </span>
+            </button>
+          </li>
+        ))}
+      </ul>
+    </section>
+  );
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/client.ts b/packages/paperclip-runner/devtools/browser/src/live/client.ts
new file mode 100644
index 00000000000..0021456c332
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/client.ts
@@ -0,0 +1,158 @@
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import type {
+  GoalOperation,
+  LiveManifestSummary,
+  LiveSessionState,
+} from "./protocol";
+
+const BASE = "/api/phase4b";
+
+export class LiveConsoleError extends Error {
+  readonly status: number;
+  readonly code: string;
+
+  constructor(status: number, code: string, message: string) {
+    super(message);
+    this.name = "LiveConsoleError";
+    this.status = status;
+    this.code = code;
+  }
+}
+
+async function send<T>(path: string, init?: RequestInit): Promise<T> {
+  const response = await fetch(`${BASE}${path}`, {
+    ...init,
+    headers: init?.body === undefined ? undefined : { "content-type": "application/json" },
+  });
+  const body = (await response.json()) as Record<string, unknown>;
+  if (!response.ok) {
+    throw new LiveConsoleError(
+      response.status,
+      typeof body.error === "string" ? body.error : "request_failed",
+      typeof body.message === "string" ? body.message : `Request to ${path} failed.`,
+    );
+  }
+  return body as T;
+}
+
+export async function fetchManifests(): Promise<LiveManifestSummary[]> {
+  const body = await send<{ manifests: LiveManifestSummary[] }>("/manifests");
+  return body.manifests;
+}
+
+export async function createSession(input: {
+  manifest: string;
+  objective: string;
+  message?: string;
+  startTurn?: boolean;
+}): Promise<LiveSessionState> {
+  return send<LiveSessionState>("/sessions", {
+    method: "POST",
+    body: JSON.stringify(input),
+  });
+}
+
+export async function readSession(sessionId: string): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}`);
+}
+
+export async function readEvents(
+  sessionId: string,
+  after = 0,
+): Promise<{ events: PrpEvent[]; cursor: number; replay: boolean }> {
+  return send(`/sessions/${sessionId}/events?after=${after}`);
+}
+
+export async function startTurn(
+  sessionId: string,
+  text: string,
+): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}/turns`, {
+    method: "POST",
+    body: JSON.stringify({ text }),
+  });
+}
+
+export async function steerTurn(
+  sessionId: string,
+  turnId: string,
+  text: string,
+): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}/steer`, {
+    method: "POST",
+    body: JSON.stringify({ turnId, text }),
+  });
+}
+
+export async function interruptTurn(
+  sessionId: string,
+  turnId: string | null,
+): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}/interrupt`, {
+    method: "POST",
+    body: JSON.stringify({
+      reason: "browser_operator",
+      ...(turnId === null ? {} : { turnId }),
+    }),
+  });
+}
+
+export async function resolveRequest(
+  sessionId: string,
+  requestId: string,
+  turnId: string,
+  resolution: Record<string, unknown>,
+): Promise<LiveSessionState> {
+  return send<LiveSessionState>(
+    `/sessions/${sessionId}/requests/${encodeURIComponent(requestId)}/resolve`,
+    { method: "POST", body: JSON.stringify({ turnId, resolution }) },
+  );
+}
+
+export async function goalOperation(
+  sessionId: string,
+  operation: GoalOperation,
+  body: Record<string, unknown> = {},
+): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}/goal/${operation}`, {
+    method: "POST",
+    body: JSON.stringify(body),
+  });
+}
+
+export async function reconnectSession(sessionId: string): Promise<LiveSessionState> {
+  return send<LiveSessionState>(`/sessions/${sessionId}/reconnect`, {
+    method: "POST",
+    body: JSON.stringify({}),
+  });
+}
+
+export interface EventStreamHandle {
+  close(): void;
+}
+
+/**
+ * Subscribes to the canonical event stream from a durable cursor. The stream
+ * never carries provider credentials; the server redacts every frame.
+ */
+export function openEventStream(
+  sessionId: string,
+  after: number,
+  handlers: {
+    onEvent: (event: PrpEvent) => void;
+    onOpen?: () => void;
+    onError?: () => void;
+  },
+): EventStreamHandle {
+  const source = new EventSource(`${BASE}/sessions/${sessionId}/stream?after=${after}`);
+  source.onopen = () => handlers.onOpen?.();
+  source.onmessage = (message) => {
+    try {
+      handlers.onEvent(JSON.parse(message.data) as PrpEvent);
+    } catch {
+      handlers.onError?.();
+    }
+  };
+  source.onerror = () => handlers.onError?.();
+  return { close: () => source.close() };
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/protocol.ts b/packages/paperclip-runner/devtools/browser/src/live/protocol.ts
new file mode 100644
index 00000000000..911f93ce161
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/protocol.ts
@@ -0,0 +1,114 @@
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import type { SessionSnapshot } from "../../../../src/reducer/session-reducer";
+
+/**
+ * The browser only ever sees the demo server's public state and the canonical
+ * PRP event stream. There is no second event model and no client-side session
+ * cache: every surface reads these shapes or the reducer snapshot they carry.
+ */
+export interface LiveCapabilities {
+  resume: boolean;
+  typedEvents: boolean;
+  steering: boolean;
+  interruption: boolean;
+  structuredResult: boolean;
+  read?: boolean;
+  reconciliation?: boolean;
+  usage?: boolean;
+  runtimeRequestResolution?: boolean;
+  goals?: boolean;
+  threadLineage?: boolean;
+  unsupported?: string[];
+}
+
+export interface LivePendingRequest {
+  requestId: string;
+  requestKind: string;
+  method: string;
+  turnId: string;
+  itemId: string;
+  status: string;
+  prompt: string;
+  details: Record<string, unknown>;
+}
+
+export interface LiveGoal {
+  threadId: string;
+  objective: string;
+  status: string;
+  tokenBudget: number | null;
+  tokensUsed: number;
+  timeUsedSeconds: number;
+  createdAt: number;
+  updatedAt: number;
+}
+
+export interface LiveLineageEntry {
+  threadId: string;
+  providerSessionId: string | null;
+  parentThreadId: string | null;
+  depth: number;
+  nickname: string | null;
+  role: string | null;
+  status: string;
+}
+
+export interface LiveSessionState {
+  sessionId: string;
+  runId: string;
+  normalizedSessionId: string;
+  manifest: string | null;
+  providerAuthentication: string;
+  credentialsExposed: boolean;
+  capabilities: LiveCapabilities;
+  driverSession: {
+    driverSessionId: string;
+    providerSessionId?: string | null;
+    displayId?: string | null;
+  };
+  activeTurnId: string | null;
+  pendingRequests: LivePendingRequest[];
+  goal: LiveGoal | null;
+  lineage: LiveLineageEntry[];
+  cursor: number;
+  snapshot: SessionSnapshot;
+}
+
+export interface LiveManifestSummary {
+  id: string;
+  name: string;
+  purpose: string;
+  scenarios: string[];
+  objective: string;
+  prompt: string;
+  expectedObservations: string[];
+}
+
+export type LiveConnectionStatus =
+  | "idle"
+  | "connecting"
+  | "connected"
+  | "reconnecting"
+  | "disconnected"
+  | "terminal";
+
+export type GoalOperation = "get" | "set" | "pause" | "resume" | "clear";
+
+export type LiveEventList = readonly PrpEvent[];
+
+export const GOAL_OPERATIONS: readonly GoalOperation[] = [
+  "set",
+  "get",
+  "pause",
+  "resume",
+  "clear",
+];
+
+/** Human labels for the five fixed goal verbs (interaction map §5). */
+export const GOAL_OPERATION_LABELS: Record<GoalOperation, string> = {
+  set: "Set goal…",
+  get: "View",
+  pause: "Pause",
+  resume: "Resume",
+  clear: "Clear",
+};
diff --git a/packages/paperclip-runner/devtools/browser/src/live/transcript-model.test.ts b/packages/paperclip-runner/devtools/browser/src/live/transcript-model.test.ts
new file mode 100644
index 00000000000..9a8458ea314
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/transcript-model.test.ts
@@ -0,0 +1,255 @@
+import { describe, expect, it } from "vitest";
+
+import { Phase4bScriptedDriver } from "../../../../src/mock-core/phase4b-scripted-driver";
+import type { HarnessSession } from "../../../../src/contracts/harness-driver";
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import {
+  applyPrpEvent,
+  createSessionSnapshotFromMetadata,
+  type SessionSnapshot,
+} from "../../../../src/reducer/session-reducer";
+import type { LiveCapabilities } from "./protocol";
+import {
+  buildTranscript,
+  capabilityRows,
+  composerState,
+  goalAvailability,
+  transcriptRole,
+  type TranscriptEntry,
+} from "./transcript-model";
+
+const FULL: LiveCapabilities = {
+  resume: true,
+  typedEvents: true,
+  steering: true,
+  interruption: true,
+  structuredResult: true,
+  runtimeRequestResolution: true,
+  goals: true,
+  threadLineage: true,
+};
+
+function seed(runId: string, sessionId: string): SessionSnapshot {
+  return createSessionSnapshotFromMetadata({
+    fixtureName: "phase4b-demo",
+    identity: {
+      schema: "paperclip.prp.identity.v1",
+      companyId: "package-local-demo",
+      issueId: "phase4b-demo",
+      runId,
+      environmentLeaseId: "package-local",
+      runnerInstanceId: "phase4b-demo-server",
+      normalizedSessionId: sessionId,
+      driverSessionId: `thread-${sessionId}`,
+    },
+    capabilities: {
+      schema: "paperclip.prp.capabilities.v1",
+      sessionReusePolicy: "reuse_per_issue",
+      driver: { kind: "phase4b-demo", version: "1" },
+      steer: true,
+      interrupt: true,
+      resume: true,
+      runtimeRequests: true,
+      structuredResult: true,
+      typedEvents: true,
+    },
+  });
+}
+
+async function drive(
+  manifestId: string,
+  interact?: (session: HarnessSession, turnId: string) => Promise<void>,
+): Promise<{ events: PrpEvent[]; snapshot: SessionSnapshot }> {
+  const driver = new Phase4bScriptedDriver({ manifestId, chunkDelayMs: 0 });
+  const session = await driver.openSession({
+    runId: `run-${manifestId}`,
+    normalizedSessionId: `session-${manifestId}`,
+    workingDirectory: "/demo",
+  });
+  const events: PrpEvent[] = [];
+  const pump = (async () => {
+    for await (const event of session.events()) events.push(event);
+  })();
+  const { turnId } = await session.startTurn({ message: { role: "user", text: "Go." } });
+  if (interact !== undefined) await interact(session, turnId);
+  await new Promise<void>((resolve) => setTimeout(resolve, 60));
+  await session.close({ reason: "test" });
+  await pump;
+  return {
+    events,
+    snapshot: events.reduce(applyPrpEvent, seed(`run-${manifestId}`, `session-${manifestId}`)),
+  };
+}
+
+function kinds(entries: readonly TranscriptEntry[]): string[] {
+  return entries.map((entry) => (entry.kind === "item" ? `item:${entry.role}` : entry.kind));
+}
+
+describe("transcriptRole", () => {
+  it("maps canonical item kinds onto transcript roles", () => {
+    expect(transcriptRole("reasoning")).toBe("reasoning");
+    expect(transcriptRole("agentMessage")).toBe("assistant");
+    expect(transcriptRole("commandExecution")).toBe("tool");
+    expect(transcriptRole("steering_acknowledgement")).toBe("steering");
+    expect(transcriptRole("thread_lineage")).toBe("lineage");
+    expect(transcriptRole("something_new")).toBe("system");
+  });
+});
+
+describe("buildTranscript", () => {
+  it("orders entries by the reducer timeline and shows the submitted message", async () => {
+    const { events, snapshot } = await drive("completion");
+    const entries = buildTranscript(snapshot, events);
+    expect(kinds(entries)).toEqual([
+      "user",
+      "item:reasoning",
+      "item:tool",
+      "item:assistant",
+      "turn",
+    ]);
+    const user = entries[0];
+    expect(user.kind === "user" && user.text).toBe("Go.");
+    const tool = entries[2];
+    expect(tool.kind === "item" && tool.input).toBe("docs/phase-04b-protocol-server.md");
+    expect(tool.kind === "item" && tool.output).toContain("server-sent events");
+  });
+
+  it("marks an interrupted item and never claims it completed cleanly", async () => {
+    const { events, snapshot } = await drive("interrupt-during-tool", async (session, turnId) => {
+      await new Promise<void>((resolve) => setTimeout(resolve, 20));
+      await session.interrupt?.({ turnId });
+    });
+    const entries = buildTranscript(snapshot, events);
+    const tool = entries.find((entry) => entry.kind === "item" && entry.role === "tool");
+    expect(tool?.kind === "item" && tool.interrupted).toBe(true);
+    expect(tool?.kind === "item" && tool.output).toBeNull();
+    const turn = entries.find((entry) => entry.kind === "turn");
+    expect(turn?.kind === "turn" && turn.state).toBe("interrupted");
+  });
+
+  it("keeps a resolved request in the record with the chosen action", async () => {
+    const { events, snapshot } = await drive("approvals", async (session, turnId) => {
+      await new Promise<void>((resolve) => setTimeout(resolve, 20));
+      await session.resolveRuntimeRequest?.({
+        requestId: "request-command",
+        turnId,
+        resolution: { action: "accept_for_session" },
+      });
+    });
+    const request = buildTranscript(snapshot, events).find(
+      (entry) => entry.kind === "request" && entry.requestId === "request-command",
+    );
+    expect(request?.kind === "request" && request.status).toBe("resolved");
+    expect(request?.kind === "request" && request.resolvedAction).toBe("accept_for_session");
+    expect(request?.kind === "request" && request.actions).toEqual([
+      "accept",
+      "accept_for_session",
+      "decline",
+      "cancel",
+    ]);
+  });
+
+  it("renders a failed item with its exact diagnostic", async () => {
+    const { events, snapshot } = await drive("failure");
+    const failed = buildTranscript(snapshot, events).find(
+      (entry) => entry.kind === "item" && entry.failure !== null,
+    );
+    expect(failed?.kind === "item" && failed.failure).toContain("cannot find module");
+  });
+
+  it("keeps a replayed prefix a real prefix of the finished transcript", async () => {
+    // Replay mode reduces events[0..n]. Stepping forward must only ever add
+    // entries in the same positions the live session produced.
+    const { events, snapshot } = await drive("completion");
+    const full = buildTranscript(snapshot, events);
+
+    for (let cut = 1; cut <= events.length; cut += 1) {
+      const prefix = events.slice(0, cut);
+      const stepped = buildTranscript(
+        prefix.reduce(applyPrpEvent, seed("run-completion", "session-completion")),
+        prefix,
+      );
+      expect(stepped.length).toBeLessThanOrEqual(full.length);
+      stepped.forEach((entry, index) => {
+        expect(entry.key).toBe(full[index]?.key);
+        expect(entry.position).toBe(full[index]?.position);
+      });
+    }
+    expect(
+      buildTranscript(
+        events.reduce(applyPrpEvent, seed("run-completion", "session-completion")),
+        events,
+      ),
+    ).toEqual(full);
+  });
+});
+
+describe("composerState", () => {
+  const base = {
+    connection: "connected" as const,
+    submitting: false,
+    interrupting: false,
+    activeTurnId: null,
+    sessionState: "running" as const,
+  };
+
+  it("shows exactly one state per protocol situation", () => {
+    expect(composerState(base)).toBe("idle");
+    expect(composerState({ ...base, submitting: true })).toBe("submitting");
+    expect(composerState({ ...base, activeTurnId: "turn-1" })).toBe("active-turn");
+    expect(composerState({ ...base, interrupting: true, activeTurnId: "turn-1" })).toBe(
+      "interrupting",
+    );
+    expect(composerState({ ...base, connection: "disconnected" })).toBe("disconnected");
+    expect(composerState({ ...base, sessionState: "closed" })).toBe("terminal");
+  });
+
+  it("puts a terminal session ahead of every other state", () => {
+    expect(
+      composerState({
+        ...base,
+        sessionState: "failed",
+        activeTurnId: "turn-1",
+        connection: "disconnected",
+      }),
+    ).toBe("terminal");
+  });
+});
+
+describe("goalAvailability", () => {
+  it("disables every verb with the exact diagnostic when goals are unsupported", () => {
+    const rows = goalAvailability(
+      {
+        ...FULL,
+        goals: false,
+        unsupported: ["Goal operations unsupported: app-server 0.132.0 does not advertise goals"],
+      },
+      null,
+    );
+    expect(rows).toHaveLength(5);
+    expect(rows.every((row) => !row.enabled)).toBe(true);
+    expect(rows[0]?.reason).toContain("does not advertise goals");
+  });
+
+  it("disables only the verbs that are invalid for the current goal state", () => {
+    const active = goalAvailability(FULL, "active");
+    expect(active.find((row) => row.operation === "pause")?.enabled).toBe(true);
+    expect(active.find((row) => row.operation === "resume")?.enabled).toBe(false);
+
+    const paused = goalAvailability(FULL, "paused");
+    expect(paused.find((row) => row.operation === "resume")?.enabled).toBe(true);
+    expect(paused.find((row) => row.operation === "pause")?.enabled).toBe(false);
+
+    const none = goalAvailability(FULL, null);
+    expect(none.find((row) => row.operation === "set")?.enabled).toBe(true);
+    expect(none.find((row) => row.operation === "clear")?.enabled).toBe(false);
+  });
+});
+
+describe("capabilityRows", () => {
+  it("gives every unsupported capability a diagnostic string", () => {
+    const rows = capabilityRows({ ...FULL, goals: false, usage: false });
+    expect(rows.filter((row) => !row.supported).every((row) => row.diagnostic !== null)).toBe(true);
+    expect(rows.find((row) => row.name === "steer")?.supported).toBe(true);
+  });
+});
diff --git a/packages/paperclip-runner/devtools/browser/src/live/transcript-model.ts b/packages/paperclip-runner/devtools/browser/src/live/transcript-model.ts
new file mode 100644
index 00000000000..6ae71f83df0
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/transcript-model.ts
@@ -0,0 +1,385 @@
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import type {
+  SessionItemSnapshot,
+  SessionSnapshot,
+} from "../../../../src/reducer/session-reducer";
+import type { GoalOperation, LiveCapabilities } from "./protocol";
+
+/**
+ * A projection of reducer state plus the canonical events that produced it.
+ * It never stores anything the protocol did not send, so a replayed session
+ * and a live session with the same events project identically.
+ */
+export type TranscriptRole =
+  | "user"
+  | "assistant"
+  | "reasoning"
+  | "tool"
+  | "steering"
+  | "interrupt"
+  | "goal"
+  | "lineage"
+  | "system";
+
+export interface TranscriptItemEntry {
+  kind: "item";
+  key: string;
+  position: number;
+  role: TranscriptRole;
+  item: SessionItemSnapshot;
+  interrupted: boolean;
+  streaming: boolean;
+  toolName: string | null;
+  input: string | null;
+  output: string | null;
+  failure: string | null;
+}
+
+export interface TranscriptUserEntry {
+  kind: "user";
+  key: string;
+  position: number;
+  turnId: string | null;
+  text: string;
+}
+
+export interface TranscriptRequestEntry {
+  kind: "request";
+  key: string;
+  position: number;
+  requestId: string;
+  requestKind: string;
+  type: string;
+  prompt: string;
+  status: "pending" | "resolved" | "expired" | "cancelled";
+  actions: string[];
+  details: Record<string, unknown>;
+  resolvedAction: string | null;
+  resolvedAt: string | null;
+  turnId: string | null;
+}
+
+export interface TranscriptTurnEntry {
+  kind: "turn";
+  key: string;
+  position: number;
+  turnId: string | null;
+  state: "completed" | "failed" | "interrupted" | "cancelled";
+  detail: string;
+}
+
+export interface TranscriptDiagnosticEntry {
+  kind: "diagnostic";
+  key: string;
+  position: number;
+  message: string;
+  code: string | null;
+}
+
+export type TranscriptEntry =
+  | TranscriptItemEntry
+  | TranscriptUserEntry
+  | TranscriptRequestEntry
+  | TranscriptTurnEntry
+  | TranscriptDiagnosticEntry;
+
+const TOOL_KINDS = new Set([
+  "commandExecution",
+  "command_execution",
+  "fileChange",
+  "file_change",
+  "mcpToolCall",
+  "tool",
+  "webSearch",
+  "patchApply",
+]);
+
+const ASSISTANT_KINDS = new Set(["agentMessage", "assistant", "assistantMessage", "message"]);
+
+export function transcriptRole(kind: string): TranscriptRole {
+  if (kind === "reasoning") return "reasoning";
+  if (ASSISTANT_KINDS.has(kind)) return "assistant";
+  if (TOOL_KINDS.has(kind)) return "tool";
+  if (kind === "steering_acknowledgement") return "steering";
+  if (kind === "interrupt_acknowledgement") return "interrupt";
+  if (kind === "goal" || kind === "goal_cleared") return "goal";
+  if (kind === "thread_lineage") return "lineage";
+  return "system";
+}
+
+function payloadRecord(value: unknown): Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : {};
+}
+
+function payloadText(value: unknown): string | null {
+  return typeof value === "string" && value.length > 0 ? value : null;
+}
+
+interface ItemFacts {
+  interrupted: boolean;
+  toolName: string | null;
+  input: string | null;
+  output: string | null;
+  failure: string | null;
+  terminal: boolean;
+}
+
+function collectItemFacts(events: readonly PrpEvent[]): Map<string, ItemFacts> {
+  const facts = new Map<string, ItemFacts>();
+  for (const event of events) {
+    if (event.itemId === undefined || !event.eventType.startsWith("item.")) continue;
+    const payload = payloadRecord(event.payload);
+    const current = facts.get(event.itemId) ?? {
+      interrupted: false,
+      toolName: null,
+      input: null,
+      output: null,
+      failure: null,
+      terminal: false,
+    };
+    if (payload.interrupted === true || payload.status === "interrupted") {
+      current.interrupted = true;
+    }
+    current.toolName = payloadText(payload.toolName) ?? current.toolName;
+    current.input = payloadText(payload.input) ?? current.input;
+    current.output = payloadText(payload.output) ?? current.output;
+    if (event.eventType === "item.failed") {
+      current.failure = payloadText(payload.message) ?? "The item failed.";
+    }
+    if (event.eventType === "item.completed" || event.eventType === "item.failed") {
+      current.terminal = true;
+    }
+    facts.set(event.itemId, current);
+  }
+  return facts;
+}
+
+function requestActions(details: unknown): string[] {
+  const record = payloadRecord(details);
+  return Array.isArray(record.actions)
+    ? record.actions.filter((action): action is string => typeof action === "string")
+    : [];
+}
+
+/**
+ * Builds the ordered transcript. Order comes from the reducer timeline, which
+ * is authoritative; nothing is re-sorted client-side.
+ */
+export function buildTranscript(
+  snapshot: SessionSnapshot,
+  events: readonly PrpEvent[],
+): TranscriptEntry[] {
+  const facts = collectItemFacts(events);
+  const items = new Map(snapshot.items.map((item) => [item.itemId, item]));
+  const seenItems = new Set<string>();
+  const entries: TranscriptEntry[] = [];
+  const eventsById = new Map(events.map((event) => [event.sourceEventId, event]));
+
+  for (const timelineEntry of snapshot.timeline) {
+    const event = eventsById.get(timelineEntry.sourceEventId);
+    const payload = payloadRecord(event?.payload);
+    const position = timelineEntry.position;
+
+    if (timelineEntry.eventType === "turn.submitted") {
+      entries.push({
+        kind: "user",
+        key: `user:${timelineEntry.sourceEventId}`,
+        position,
+        turnId: typeof payload.turnId === "string" ? payload.turnId : null,
+        text: payloadText(payload.text) ?? "(no message text in the canonical event)",
+      });
+      continue;
+    }
+
+    if (
+      ["turn.completed", "turn.failed", "turn.interrupted", "turn.cancelled"].includes(
+        timelineEntry.eventType,
+      )
+    ) {
+      const state = timelineEntry.eventType.split(".").at(-1) as TranscriptTurnEntry["state"];
+      entries.push({
+        kind: "turn",
+        key: `turn:${timelineEntry.sourceEventId}`,
+        position,
+        turnId: typeof payload.turnId === "string" ? payload.turnId : null,
+        state,
+        detail:
+          payloadText(payload.message) ??
+          payloadText(payload.reason) ??
+          timelineEntry.summary,
+      });
+      continue;
+    }
+
+    if (timelineEntry.eventType === "runtime_request.created") {
+      const request = payloadRecord(payload.request ?? payload);
+      const requestId = typeof request.requestId === "string" ? request.requestId : "";
+      const projected = snapshot.requests.find(
+        (candidate) => candidate.requestId === requestId,
+      );
+      const resolution = events.find(
+        (candidate) =>
+          candidate.eventType.startsWith("runtime_request.") &&
+          candidate.eventType !== "runtime_request.created" &&
+          payloadRecord(candidate.payload).requestId === requestId,
+      );
+      entries.push({
+        kind: "request",
+        key: `request:${requestId}`,
+        position,
+        requestId,
+        requestKind: typeof request.requestKind === "string" ? request.requestKind : "runtime",
+        type: typeof request.type === "string" ? request.type : "runtime",
+        prompt: payloadText(request.prompt) ?? "Runtime request",
+        status: (projected?.status ?? "pending") as TranscriptRequestEntry["status"],
+        actions: requestActions(request),
+        details: payloadRecord(request.details),
+        resolvedAction: payloadText(payloadRecord(resolution?.payload).action),
+        resolvedAt: resolution?.emittedAt ?? null,
+        turnId: typeof request.turnId === "string" ? request.turnId : null,
+      });
+      continue;
+    }
+
+    if (
+      timelineEntry.eventType === "harness.diagnostic" ||
+      timelineEntry.eventType === "runner.diagnostic"
+    ) {
+      entries.push({
+        kind: "diagnostic",
+        key: `diagnostic:${timelineEntry.sourceEventId}`,
+        position,
+        message: payloadText(payload.message) ?? timelineEntry.summary,
+        code: payloadText(payload.code),
+      });
+      continue;
+    }
+
+    const itemId = timelineEntry.itemId;
+    if (itemId === undefined || seenItems.has(itemId)) continue;
+    const item = items.get(itemId);
+    if (item === undefined) continue;
+    seenItems.add(itemId);
+    const itemFacts = facts.get(itemId);
+    entries.push({
+      kind: "item",
+      key: `item:${itemId}`,
+      position,
+      role: transcriptRole(item.kind),
+      item,
+      interrupted: itemFacts?.interrupted ?? false,
+      streaming: item.status === "running" && itemFacts?.terminal !== true,
+      toolName: itemFacts?.toolName ?? null,
+      input: itemFacts?.input ?? null,
+      output: itemFacts?.output ?? null,
+      failure: itemFacts?.failure ?? null,
+    });
+  }
+
+  return entries;
+}
+
+export type ComposerState =
+  | "idle"
+  | "submitting"
+  | "active-turn"
+  | "interrupting"
+  | "disconnected"
+  | "terminal";
+
+export interface ComposerStateInput {
+  connection: "idle" | "connecting" | "connected" | "reconnecting" | "disconnected" | "terminal";
+  submitting: boolean;
+  interrupting: boolean;
+  activeTurnId: string | null;
+  sessionState: SessionSnapshot["sessionState"];
+}
+
+/** One state machine with exactly one visible primary action. */
+export function composerState(input: ComposerStateInput): ComposerState {
+  if (input.sessionState === "closed" || input.sessionState === "failed") return "terminal";
+  if (input.connection === "disconnected" || input.connection === "reconnecting") {
+    return "disconnected";
+  }
+  if (input.interrupting) return "interrupting";
+  if (input.activeTurnId !== null) return "active-turn";
+  if (input.submitting) return "submitting";
+  return "idle";
+}
+
+export interface GoalOperationAvailability {
+  operation: GoalOperation;
+  enabled: boolean;
+  reason: string | null;
+}
+
+const GOAL_UNSUPPORTED_FALLBACK =
+  "Goal operations unsupported: this app-server does not advertise the goals capability";
+
+/**
+ * Capability gating for the goal menu. An unsupported verb is disabled with
+ * the exact upstream diagnostic — never hidden and never emulated.
+ */
+export function goalAvailability(
+  capabilities: LiveCapabilities,
+  goalStatus: string | null,
+): GoalOperationAvailability[] {
+  const unsupported = capabilities.unsupported ?? [];
+  if (capabilities.goals !== true) {
+    const reason = unsupported[0] ?? GOAL_UNSUPPORTED_FALLBACK;
+    return (["set", "get", "pause", "resume", "clear"] as GoalOperation[]).map((operation) => ({
+      operation,
+      enabled: false,
+      reason,
+    }));
+  }
+  return (["set", "get", "pause", "resume", "clear"] as GoalOperation[]).map((operation) => {
+    if (operation === "set") return { operation, enabled: true, reason: null };
+    if (goalStatus === null) {
+      return {
+        operation,
+        enabled: false,
+        reason: "No goal is set on this thread yet.",
+      };
+    }
+    if (operation === "pause" && goalStatus !== "active") {
+      return { operation, enabled: false, reason: `The goal is already ${goalStatus}.` };
+    }
+    if (operation === "resume" && goalStatus === "active") {
+      return { operation, enabled: false, reason: "The goal is already active." };
+    }
+    return { operation, enabled: true, reason: null };
+  });
+}
+
+export interface CapabilityRow {
+  name: string;
+  supported: boolean;
+  diagnostic: string | null;
+}
+
+export function capabilityRows(capabilities: LiveCapabilities): CapabilityRow[] {
+  const unsupported = capabilities.unsupported ?? [];
+  const rows: Array<[string, boolean | undefined]> = [
+    ["steer", capabilities.steering],
+    ["interrupt", capabilities.interruption],
+    ["resume", capabilities.resume],
+    ["runtime requests", capabilities.runtimeRequestResolution],
+    ["goals", capabilities.goals],
+    ["child threads", capabilities.threadLineage],
+    ["structured result", capabilities.structuredResult],
+    ["typed events", capabilities.typedEvents],
+    ["usage", capabilities.usage],
+  ];
+  return rows.map(([name, supported]) => ({
+    name,
+    supported: supported === true,
+    diagnostic:
+      supported === true
+        ? null
+        : (unsupported.find((entry) => entry.toLowerCase().includes(name.split(" ")[0]!)) ??
+          `${name} is not advertised by this app-server`),
+  }));
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/live/use-live-console.ts b/packages/paperclip-runner/devtools/browser/src/live/use-live-console.ts
new file mode 100644
index 00000000000..66c678edee3
--- /dev/null
+++ b/packages/paperclip-runner/devtools/browser/src/live/use-live-console.ts
@@ -0,0 +1,550 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+
+import type { PrpEvent } from "../../../../src/protocol/phase1-contract";
+import {
+  applyPrpEvent,
+  createSessionSnapshotFromMetadata,
+  type SessionSnapshot,
+} from "../../../../src/reducer/session-reducer";
+import * as client from "./client";
+import type {
+  GoalOperation,
+  LiveConnectionStatus,
+  LiveManifestSummary,
+  LiveSessionState,
+} from "./protocol";
+import {
+  buildTranscript,
+  composerState,
+  type ComposerState,
+  type TranscriptEntry,
+} from "./transcript-model";
+
+const SESSION_STORAGE_KEY = "paperclip-runner.phase4b.session";
+const RECONNECT_DELAY_MS = 900;
+const MAX_RECONNECT_ATTEMPTS = 5;
+
+export type SteeringChipStatus = "pending" | "acknowledged" | "rejected" | "failed";
+
+export interface SteeringChip {
+  id: string;
+  text: string;
+  expectedTurnId: string;
+  status: SteeringChipStatus;
+  detail: string | null;
+}
+
+function reduceEvents(
+  state: LiveSessionState | null,
+  events: readonly PrpEvent[],
+): SessionSnapshot | null {
+  if (state === null) return null;
+  const seed = createSessionSnapshotFromMetadata({
+    fixtureName: state.snapshot.fixtureName,
+    identity: state.snapshot.identity,
+    capabilities: state.snapshot.capabilities,
+  });
+  return events.reduce(applyPrpEvent, seed);
+}
+
+function errorMessage(cause: unknown): string {
+  if (cause instanceof client.LiveConsoleError) return cause.message;
+  return cause instanceof Error ? cause.message : String(cause);
+}
+
+export interface LiveConsole {
+  manifests: LiveManifestSummary[];
+  selectedManifestId: string;
+  selectManifest: (id: string) => void;
+  state: LiveSessionState | null;
+  events: PrpEvent[];
+  snapshot: SessionSnapshot | null;
+  transcript: TranscriptEntry[];
+  connection: LiveConnectionStatus;
+  reconnectAttempt: number;
+  replayParity: boolean | null;
+  composer: ComposerState;
+  steeringChips: SteeringChip[];
+  announcement: string;
+  error: string | null;
+  busy: boolean;
+  selectedThreadId: string | null;
+  selectThread: (threadId: string | null) => void;
+  replay: {
+    active: boolean;
+    position: number;
+    length: number;
+    playing: boolean;
+    enter: () => void;
+    exit: () => void;
+    setPosition: (position: number) => void;
+    togglePlay: () => void;
+    step: (delta: number) => void;
+  };
+  start: (input: { manifestId: string; message: string }) => Promise<void>;
+  send: (text: string) => Promise<void>;
+  steer: (text: string) => Promise<void>;
+  interrupt: () => Promise<void>;
+  resolve: (
+    requestId: string,
+    turnId: string,
+    resolution: Record<string, unknown>,
+  ) => Promise<void>;
+  goal: (operation: GoalOperation, body?: Record<string, unknown>) => Promise<void>;
+  dropConnection: () => void;
+  retryNow: () => void;
+  reset: () => Promise<void>;
+  dismissSteeringChip: (id: string) => void;
+}
+
+export function useLiveConsole(): LiveConsole {
+  const [manifests, setManifests] = useState<LiveManifestSummary[]>([]);
+  const [selectedManifestId, setSelectedManifestId] = useState("completion");
+  const [state, setState] = useState<LiveSessionState | null>(null);
+  const [events, setEvents] = useState<PrpEvent[]>([]);
+  const [connection, setConnection] = useState<LiveConnectionStatus>("idle");
+  const [reconnectAttempt, setReconnectAttempt] = useState(0);
+  const [steeringChips, setSteeringChips] = useState<SteeringChip[]>([]);
+  const [announcement, setAnnouncement] = useState("");
+  const [error, setError] = useState<string | null>(null);
+  const [busy, setBusy] = useState(false);
+  const [interrupting, setInterrupting] = useState(false);
+  const [submitting, setSubmitting] = useState(false);
+  const [selectedThreadId, setSelectedThreadId] = useState<string | null>(null);
+  const [replayActive, setReplayActive] = useState(false);
+  const [replayPosition, setReplayPosition] = useState(0);
+  const [replayPlaying, setReplayPlaying] = useState(false);
+
+  const streamRef = useRef<client.EventStreamHandle | null>(null);
+  const cursorRef = useRef(0);
+  const sessionIdRef = useRef<string | null>(null);
+  const retryTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+
+  const closeStream = useCallback(() => {
+    streamRef.current?.close();
+    streamRef.current = null;
+    if (retryTimerRef.current !== null) {
+      clearTimeout(retryTimerRef.current);
+      retryTimerRef.current = null;
+    }
+  }, []);
+
+  const subscribe = useCallback(
+    (sessionId: string, attempt = 0) => {
+      closeStream();
+      setConnection(attempt === 0 ? "connecting" : "reconnecting");
+      setReconnectAttempt(attempt);
+      streamRef.current = client.openEventStream(sessionId, cursorRef.current, {
+        onOpen: () => {
+          setConnection("connected");
+          setReconnectAttempt(0);
+          setAnnouncement("Connected");
+        },
+        onEvent: (event) => {
+          cursorRef.current += 1;
+          setEvents((current) =>
+            current.some((candidate) => candidate.sourceEventId === event.sourceEventId)
+              ? current
+              : [...current, event],
+          );
+        },
+        onError: () => {
+          closeStream();
+          if (attempt >= MAX_RECONNECT_ATTEMPTS) {
+            setConnection("disconnected");
+            setAnnouncement("Disconnected");
+            return;
+          }
+          setConnection("reconnecting");
+          setReconnectAttempt(attempt + 1);
+          setAnnouncement(`Reconnecting, attempt ${attempt + 1}`);
+          retryTimerRef.current = setTimeout(
+            () => subscribe(sessionId, attempt + 1),
+            RECONNECT_DELAY_MS,
+          );
+        },
+      });
+    },
+    [closeStream],
+  );
+
+  const adopt = useCallback(
+    (next: LiveSessionState) => {
+      setState(next);
+      setError(null);
+    },
+    [],
+  );
+
+  const attach = useCallback(
+    async (sessionId: string, resume: boolean) => {
+      const [next, history] = await Promise.all([
+        client.readSession(sessionId),
+        client.readEvents(sessionId, 0),
+      ]);
+      sessionIdRef.current = sessionId;
+      window.sessionStorage.setItem(SESSION_STORAGE_KEY, sessionId);
+      adopt(next);
+      setEvents(history.events);
+      cursorRef.current = history.cursor;
+      if (resume) setAnnouncement("Replayed the durable transcript");
+      subscribe(sessionId, 0);
+    },
+    [adopt, subscribe],
+  );
+
+  useEffect(() => {
+    let cancelled = false;
+    void client
+      .fetchManifests()
+      .then((list) => {
+        if (cancelled) return;
+        setManifests(list);
+        if (list[0] !== undefined) setSelectedManifestId((current) => current || list[0]!.id);
+      })
+      .catch((cause) => {
+        if (!cancelled) setError(errorMessage(cause));
+      });
+    const stored = window.sessionStorage.getItem(SESSION_STORAGE_KEY);
+    if (stored !== null) {
+      void attach(stored, true).catch(() => {
+        window.sessionStorage.removeItem(SESSION_STORAGE_KEY);
+      });
+    }
+    return () => {
+      cancelled = true;
+    };
+    // Runs once: session restoration is a mount concern.
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []);
+
+  useEffect(() => () => closeStream(), [closeStream]);
+
+  // Lineage, goal, pending requests, and the server cursor live in the public
+  // state, so a stream that only carries events would leave them stale.
+  useEffect(() => {
+    const sessionId = sessionIdRef.current;
+    if (sessionId === null || events.length === 0) return undefined;
+    const timer = setTimeout(() => {
+      void client
+        .readSession(sessionId)
+        .then((next) => setState(next))
+        .catch(() => undefined);
+    }, 150);
+    return () => clearTimeout(timer);
+  }, [events.length]);
+
+  const liveSnapshot = useMemo(() => reduceEvents(state, events), [state, events]);
+  const visibleEvents = useMemo(
+    () => (replayActive ? events.slice(0, replayPosition) : events),
+    [events, replayActive, replayPosition],
+  );
+  const snapshot = useMemo(
+    () => (replayActive ? reduceEvents(state, visibleEvents) : liveSnapshot),
+    [replayActive, state, visibleEvents, liveSnapshot],
+  );
+
+  const replayParity = useMemo(() => {
+    if (state === null || liveSnapshot === null) return null;
+    if (state.cursor !== events.length) return null;
+    return (
+      JSON.stringify(liveSnapshot.timeline) === JSON.stringify(state.snapshot.timeline) &&
+      liveSnapshot.integrity === state.snapshot.integrity
+    );
+  }, [state, liveSnapshot, events.length]);
+
+  const transcript = useMemo(
+    () => (snapshot === null ? [] : buildTranscript(snapshot, visibleEvents)),
+    [snapshot, visibleEvents],
+  );
+
+  useEffect(() => {
+    if (!replayPlaying) return undefined;
+    if (replayPosition >= events.length) {
+      setReplayPlaying(false);
+      return undefined;
+    }
+    const timer = setTimeout(() => setReplayPosition((current) => current + 1), 140);
+    return () => clearTimeout(timer);
+  }, [replayPlaying, replayPosition, events.length]);
+
+  // Steering chips resolve against canonical acknowledgements only.
+  useEffect(() => {
+    setSteeringChips((chips) => {
+      if (chips.every((chip) => chip.status !== "pending")) return chips;
+      const acknowledged = events.filter(
+        (event) =>
+          event.eventType === "item.completed" &&
+          (event.payload as { kind?: string }).kind === "steering_acknowledgement",
+      );
+      let index = 0;
+      return chips.map((chip) => {
+        if (chip.status !== "pending") return chip;
+        const match = acknowledged[index];
+        index += 1;
+        return match === undefined
+          ? chip
+          : { ...chip, status: "acknowledged" as const, detail: null };
+      });
+    });
+  }, [events]);
+
+  // The reducer is authoritative for the active turn. Falling back to the
+  // last public-state read would keep a finished turn alive in the composer.
+  const activeTurnId = snapshot === null ? (state?.activeTurnId ?? null) : snapshot.activeTurnId;
+
+  // A submitted turn that has not been accepted yet is a protocol fact, not a
+  // local flag: it is what makes interrupt-before-start reachable.
+  const awaitingTurnStart = useMemo(() => {
+    let pending = false;
+    for (const entry of snapshot?.timeline ?? []) {
+      if (entry.eventType === "turn.submitted") pending = true;
+      else if (entry.eventType.startsWith("turn.")) pending = false;
+    }
+    return pending;
+  }, [snapshot]);
+
+  const composer = composerState({
+    connection,
+    submitting: submitting || awaitingTurnStart,
+    interrupting,
+    activeTurnId,
+    sessionState: snapshot?.sessionState ?? "not_started",
+  });
+
+  const guard = useCallback(async (task: () => Promise<void>) => {
+    setBusy(true);
+    try {
+      await task();
+    } catch (cause) {
+      setError(errorMessage(cause));
+    } finally {
+      setBusy(false);
+    }
+  }, []);
+
+  const start = useCallback(
+    async (input: { manifestId: string; message: string }) => {
+      await guard(async () => {
+        closeStream();
+        setEvents([]);
+        cursorRef.current = 0;
+        setSteeringChips([]);
+        setReplayActive(false);
+        setReplayPosition(0);
+        setSelectedThreadId(null);
+        setSubmitting(true);
+        const manifest = manifests.find((entry) => entry.id === input.manifestId);
+        try {
+          const created = await client.createSession({
+            manifest: input.manifestId,
+            objective: manifest?.objective ?? "Run the Phase 4b demo.",
+            message: input.message,
+          });
+          sessionIdRef.current = created.sessionId;
+          window.sessionStorage.setItem(SESSION_STORAGE_KEY, created.sessionId);
+          adopt(created);
+          setAnnouncement("Turn running");
+          subscribe(created.sessionId, 0);
+        } finally {
+          setSubmitting(false);
+        }
+      });
+    },
+    [adopt, closeStream, guard, manifests, subscribe],
+  );
+
+  const refreshState = useCallback(async () => {
+    const sessionId = sessionIdRef.current;
+    if (sessionId === null) return;
+    adopt(await client.readSession(sessionId));
+  }, [adopt]);
+
+  const send = useCallback(
+    async (text: string) => {
+      const sessionId = sessionIdRef.current;
+      if (sessionId === null) {
+        await start({ manifestId: selectedManifestId, message: text });
+        return;
+      }
+      await guard(async () => {
+        setSubmitting(true);
+        try {
+          adopt(await client.startTurn(sessionId, text));
+          setAnnouncement("Turn running");
+        } finally {
+          setSubmitting(false);
+        }
+      });
+    },
+    [adopt, guard, selectedManifestId, start],
+  );
+
+  const steer = useCallback(
+    async (text: string) => {
+      const sessionId = sessionIdRef.current;
+      const expectedTurnId = activeTurnId;
+      if (sessionId === null || expectedTurnId === null) return;
+      const chipId = `steer-${Date.now()}-${steeringChips.length}`;
+      setSteeringChips((chips) => [
+        ...chips,
+        { id: chipId, text, expectedTurnId, status: "pending", detail: null },
+      ]);
+      try {
+        adopt(await client.steerTurn(sessionId, expectedTurnId, text));
+        setAnnouncement("Steering sent");
+      } catch (cause) {
+        const stale =
+          cause instanceof client.LiveConsoleError &&
+          (cause.code === "stale_turn" || cause.code === "already_terminal");
+        setSteeringChips((chips) =>
+          chips.map((chip) =>
+            chip.id === chipId
+              ? {
+                  ...chip,
+                  status: stale ? "rejected" : "failed",
+                  detail: stale ? "turn already ended" : errorMessage(cause),
+                }
+              : chip,
+          ),
+        );
+        setAnnouncement(stale ? "Steering rejected: turn already ended" : "Steering failed");
+        await refreshState().catch(() => undefined);
+      }
+    },
+    [activeTurnId, adopt, refreshState, steeringChips.length],
+  );
+
+  const interrupt = useCallback(async () => {
+    const sessionId = sessionIdRef.current;
+    if (sessionId === null) return;
+    setInterrupting(true);
+    setAnnouncement("Stopping");
+    try {
+      adopt(await client.interruptTurn(sessionId, activeTurnId));
+      setAnnouncement("Turn stopped");
+    } catch (cause) {
+      setError(errorMessage(cause));
+    } finally {
+      setInterrupting(false);
+    }
+  }, [activeTurnId, adopt]);
+
+  const resolve = useCallback(
+    async (requestId: string, turnId: string, resolution: Record<string, unknown>) => {
+      const sessionId = sessionIdRef.current;
+      if (sessionId === null) return;
+      await guard(async () => {
+        adopt(await client.resolveRequest(sessionId, requestId, turnId, resolution));
+        setAnnouncement("Request resolved");
+      });
+    },
+    [adopt, guard],
+  );
+
+  const goal = useCallback(
+    async (operation: GoalOperation, body: Record<string, unknown> = {}) => {
+      const sessionId = sessionIdRef.current;
+      if (sessionId === null) return;
+      await guard(async () => {
+        adopt(await client.goalOperation(sessionId, operation, body));
+        setAnnouncement(`Goal ${operation}`);
+      });
+    },
+    [adopt, guard],
+  );
+
+  const dropConnection = useCallback(() => {
+    closeStream();
+    setConnection("disconnected");
+    setAnnouncement("Connection lost");
+  }, [closeStream]);
+
+  const retryNow = useCallback(() => {
+    const sessionId = sessionIdRef.current;
+    if (sessionId === null) return;
+    void guard(async () => {
+      const resumed = await client.reconnectSession(sessionId);
+      adopt(resumed);
+      const history = await client.readEvents(sessionId, 0);
+      setEvents(history.events);
+      cursorRef.current = history.cursor;
+      subscribe(sessionId, 0);
+      setAnnouncement("Resumed the same session");
+    });
+  }, [adopt, guard, subscribe]);
+
+  const reset = useCallback(async () => {
+    closeStream();
+    window.sessionStorage.removeItem(SESSION_STORAGE_KEY);
+    sessionIdRef.current = null;
+    cursorRef.current = 0;
+    setState(null);
+    setEvents([]);
+    setSteeringChips([]);
+    setConnection("idle");
+    setReplayActive(false);
+    setReplayPosition(0);
+    setSelectedThreadId(null);
+    setError(null);
+    setAnnouncement("Demo state reset");
+  }, [closeStream]);
+
+  const replay = useMemo(
+    () => ({
+      active: replayActive,
+      position: replayPosition,
+      length: events.length,
+      playing: replayPlaying,
+      enter: () => {
+        setReplayActive(true);
+        setReplayPosition(0);
+        setReplayPlaying(false);
+      },
+      exit: () => {
+        setReplayActive(false);
+        setReplayPlaying(false);
+      },
+      setPosition: (position: number) =>
+        setReplayPosition(Math.min(Math.max(position, 0), events.length)),
+      togglePlay: () => setReplayPlaying((current) => !current),
+      step: (delta: number) =>
+        setReplayPosition((current) =>
+          Math.min(Math.max(current + delta, 0), events.length),
+        ),
+    }),
+    [events.length, replayActive, replayPlaying, replayPosition],
+  );
+
+  return {
+    manifests,
+    selectedManifestId,
+    selectManifest: setSelectedManifestId,
+    state,
+    events,
+    snapshot,
+    transcript,
+    connection,
+    reconnectAttempt,
+    replayParity,
+    composer,
+    steeringChips,
+    announcement,
+    error,
+    busy,
+    selectedThreadId,
+    selectThread: setSelectedThreadId,
+    replay,
+    start,
+    send,
+    steer,
+    interrupt,
+    resolve,
+    goal,
+    dropConnection,
+    retryNow,
+    reset,
+    dismissSteeringChip: (id: string) =>
+      setSteeringChips((chips) => chips.filter((chip) => chip.id !== id)),
+  };
+}
diff --git a/packages/paperclip-runner/devtools/browser/src/styles.css b/packages/paperclip-runner/devtools/browser/src/styles.css
index 34a8580dc16..422b24418ee 100644
--- a/packages/paperclip-runner/devtools/browser/src/styles.css
+++ b/packages/paperclip-runner/devtools/browser/src/styles.css
@@ -15,6 +15,10 @@
   --warning-surface: #fff4dc;
   --danger: #a02f2f;
   --danger-surface: #fceaea;
+  /* Pending and streaming are not warnings; Phase 4b needs their own pair. */
+  --accent: #2f5fa8;
+  --accent-surface: #e9f0fa;
+  --surface-raised: #fcfcfa;
   --ring: #66665f;
   --font-sans: Inter, ui-sans-serif, system-ui, sans-serif;
   --font-mono: "SFMono-Regular", Consolas, "Liberation Mono", monospace;
@@ -43,6 +47,15 @@
   --control-height: 2.5rem;
   --timeline-column: 2rem;
   --motion-fast: 120ms;
+  --motion-medium: 200ms;
+  --z-banner: 10;
+  --z-dialog: 20;
+  --rail-width: 17rem;
+  --center-width: 48rem;
+  --inspector-width: 24rem;
+  --transcript-height: 34rem;
+  --touch-target: 2.75rem;
+  --facet-height: 6rem;
 }
 
 * {
@@ -545,3 +558,910 @@ dd {
     --motion-fast: 0ms;
   }
 }
+
+/* ---------------------------------------------------------------------------
+   Phase 4b live console
+   Every value below comes from the token layer above. Adapted shadcn/ui and
+   AI Elements components carry no visual values of their own.
+   --------------------------------------------------------------------------- */
+
+.ui-badge[data-tone="accent"] {
+  background: var(--accent-surface);
+  color: var(--accent);
+}
+
+.ui-button--danger {
+  background: var(--danger);
+  color: var(--primary-foreground);
+}
+
+.ui-button--secondary {
+  background: var(--muted);
+  color: var(--foreground);
+}
+
+.ui-button--quiet {
+  background: transparent;
+  color: var(--foreground);
+  border: 0.0625rem solid var(--border);
+}
+
+.ui-visually-hidden {
+  position: absolute;
+  width: 0.0625rem;
+  height: 0.0625rem;
+  margin: -0.0625rem;
+  padding: 0;
+  overflow: hidden;
+  clip-path: inset(50%);
+  white-space: nowrap;
+}
+
+.live-console {
+  display: grid;
+  gap: var(--space-4);
+}
+
+.live-console-header,
+.live-console-title,
+.live-console-header-actions {
+  display: flex;
+  align-items: center;
+  gap: var(--space-3);
+}
+
+.live-console-header {
+  justify-content: space-between;
+  flex-wrap: wrap;
+}
+
+.live-console-title h2 {
+  font-size: var(--font-size-lg);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-breadcrumb {
+  color: var(--muted-foreground);
+  font-family: var(--font-mono);
+  font-size: var(--font-size-xs);
+}
+
+.live-console-desktop {
+  display: grid;
+  grid-template-columns: var(--rail-width) minmax(0, 1fr);
+  gap: var(--space-5);
+  align-items: start;
+}
+
+.live-console[data-inspector="true"] .live-console-desktop {
+  grid-template-columns: var(--rail-width) minmax(0, 1fr) var(--inspector-width);
+}
+
+.live-console-rail,
+.live-console-inspector {
+  display: grid;
+  /* An auto track would size to the widest identifier and overflow the rail. */
+  grid-template-columns: minmax(0, 1fr);
+  gap: var(--space-4);
+  min-width: 0;
+}
+
+.live-console-center {
+  display: grid;
+  gap: var(--space-3);
+  min-width: 0;
+  max-width: var(--center-width);
+}
+
+.ui-panel {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr);
+  min-width: 0;
+  gap: var(--space-3);
+  padding: var(--space-4);
+  background: var(--card);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+}
+
+.ui-panel-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: var(--space-2);
+}
+
+.ui-panel-title {
+  font-size: var(--font-size-sm);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-panel-actions {
+  display: grid;
+  gap: var(--space-2);
+}
+
+.ui-status-row {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: var(--space-2);
+}
+
+.ui-status-row > span:first-child {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-identity-grid {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr);
+  gap: var(--space-2);
+  margin: 0;
+  min-width: 0;
+}
+
+.ui-identity-grid > div,
+.ui-inspector-grid > div,
+.ui-status-row,
+.ui-tree-row,
+.ui-capability-list li {
+  min-width: 0;
+}
+
+.ui-identity-grid dd,
+.ui-inspector-grid dd {
+  overflow-wrap: anywhere;
+  white-space: normal;
+}
+
+/* Manifest list ---------------------------------------------------------- */
+
+.ui-manifest-list,
+.ui-tree,
+.ui-transcript,
+.ui-inspector-list,
+.ui-observation-list {
+  display: grid;
+  gap: var(--space-2);
+  margin: 0;
+  padding: 0;
+  list-style: none;
+}
+
+.ui-manifest-row {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr);
+  gap: var(--space-1);
+  width: 100%;
+  min-width: 0;
+  min-height: var(--touch-target);
+  padding: var(--space-2);
+  border: 0.0625rem solid transparent;
+  border-radius: var(--radius-sm);
+  background: transparent;
+  color: inherit;
+  text-align: left;
+  cursor: pointer;
+}
+
+.ui-manifest-row[aria-pressed="true"] {
+  border-color: var(--accent);
+  background: var(--accent-surface);
+}
+
+.ui-manifest-name {
+  font-weight: var(--weight-semibold);
+}
+
+.ui-manifest-purpose {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+  overflow-wrap: anywhere;
+}
+
+.ui-manifest-badges {
+  display: flex;
+  flex-wrap: wrap;
+  gap: var(--space-1);
+}
+
+.ui-observation-list {
+  gap: var(--space-1);
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+}
+
+/* Lineage tree ----------------------------------------------------------- */
+
+.ui-tree-row {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: var(--space-2);
+  width: 100%;
+  min-height: var(--touch-target);
+  padding: var(--space-2);
+  border: 0.0625rem solid transparent;
+  border-radius: var(--radius-sm);
+  background: transparent;
+  color: inherit;
+  cursor: pointer;
+}
+
+.ui-tree li[data-depth="1"] .ui-tree-row {
+  padding-left: var(--space-5);
+}
+
+.ui-tree-row[aria-current="true"] {
+  border-color: var(--border);
+  background: var(--muted);
+}
+
+.ui-tree-state,
+.ui-tree-empty {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+}
+
+.ui-state-dot {
+  display: inline-block;
+  width: var(--space-2);
+  height: var(--space-2);
+  border-radius: 999rem;
+  background: var(--muted-foreground);
+}
+
+.ui-state-dot[data-state="running"] {
+  background: var(--accent);
+}
+
+.ui-state-dot[data-state="completed"] {
+  background: var(--success);
+}
+
+.ui-state-dot[data-state="failed"] {
+  background: var(--danger);
+}
+
+/* Conversation ----------------------------------------------------------- */
+
+.ui-conversation {
+  position: relative;
+}
+
+.ui-conversation-viewport {
+  max-height: var(--transcript-height);
+  overflow-y: auto;
+  padding: var(--space-4);
+  background: var(--card);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+}
+
+.ui-jump-to-latest {
+  position: absolute;
+  bottom: var(--space-3);
+  left: 50%;
+  translate: -50% 0;
+  z-index: var(--z-banner);
+  min-height: var(--touch-target);
+  padding-inline: var(--space-4);
+  border: 0.0625rem solid var(--border);
+  border-radius: 999rem;
+  background: var(--card);
+  color: var(--foreground);
+  cursor: pointer;
+}
+
+.ui-empty-transcript {
+  display: grid;
+  gap: var(--space-2);
+  padding-block: var(--space-6);
+}
+
+.ui-empty-transcript h3 {
+  font-size: var(--font-size-md);
+}
+
+.ui-empty-transcript p:last-child {
+  color: var(--muted-foreground);
+}
+
+.ui-transcript {
+  gap: var(--space-3);
+}
+
+.ui-message {
+  display: grid;
+  gap: var(--space-1);
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--surface-raised);
+}
+
+.ui-message[data-role="user"] {
+  margin-left: var(--space-10);
+  background: var(--muted);
+}
+
+.ui-message[data-state="failed"] {
+  border-color: var(--danger);
+  background: var(--danger-surface);
+  color: var(--danger);
+}
+
+.ui-message[data-state="interrupted"] {
+  border-color: var(--warning);
+}
+
+.ui-message-role {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+  font-weight: var(--weight-semibold);
+  letter-spacing: 0.08em;
+  text-transform: uppercase;
+}
+
+.ui-message-text {
+  white-space: pre-wrap;
+}
+
+.ui-message-divider {
+  padding-top: var(--space-2);
+  border-top: 0.0625rem solid var(--warning);
+  color: var(--warning);
+  font-size: var(--font-size-xs);
+}
+
+.ui-stream-cursor {
+  display: inline-block;
+  width: var(--space-2);
+  height: var(--font-size-md);
+  margin-left: var(--space-1);
+  background: var(--accent);
+  vertical-align: text-bottom;
+  animation: ui-pulse 1s steps(2, end) infinite;
+}
+
+@keyframes ui-pulse {
+  50% {
+    opacity: 0.2;
+  }
+}
+
+.ui-turn-marker {
+  display: flex;
+  align-items: center;
+  gap: var(--space-2);
+  padding-block: var(--space-2);
+  border-top: 0.0625rem solid var(--border);
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+}
+
+.ui-steering-chip {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: var(--space-2);
+  padding: var(--space-2);
+  border: 0.0625rem dashed var(--border);
+  border-radius: var(--radius-sm);
+  font-size: var(--font-size-xs);
+}
+
+/* Disclosures ------------------------------------------------------------ */
+
+.ui-disclosure {
+  padding: var(--space-2) var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--surface-raised);
+}
+
+.ui-disclosure-summary {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: var(--space-2);
+  min-height: var(--touch-target);
+  cursor: pointer;
+  font-weight: var(--weight-medium);
+}
+
+.ui-disclosure-preview,
+.ui-disclosure-label {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-disclosure-preview {
+  overflow: hidden;
+  min-width: 0;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.ui-payload {
+  max-height: var(--transcript-height);
+  margin: var(--space-2) 0 0;
+  overflow: auto;
+  padding: var(--space-3);
+  border-radius: var(--radius-sm);
+  background: var(--muted);
+  font-family: var(--font-mono);
+  font-size: var(--font-size-xs);
+  white-space: pre-wrap;
+  word-break: break-word;
+}
+
+.ui-payload--danger {
+  background: var(--danger-surface);
+  color: var(--danger);
+}
+
+/* Request cards ---------------------------------------------------------- */
+
+.ui-request {
+  display: grid;
+  gap: var(--space-2);
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-left: var(--space-1) solid var(--accent);
+  border-radius: var(--radius-md);
+  background: var(--card);
+}
+
+.ui-request--settled {
+  border-left-color: var(--border);
+  background: var(--surface-raised);
+}
+
+.ui-request[data-status="resolved"] {
+  border-left-color: var(--success);
+}
+
+.ui-request-header,
+.ui-request-summary,
+.ui-request-actions,
+.ui-request-meta {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: var(--space-2);
+}
+
+.ui-request-summary {
+  justify-content: space-between;
+  min-height: var(--touch-target);
+  cursor: pointer;
+}
+
+.ui-request-kind {
+  font-weight: var(--weight-semibold);
+}
+
+.ui-request-meta,
+.ui-request-resolving {
+  color: var(--muted-foreground);
+  font-family: var(--font-mono);
+  font-size: var(--font-size-xs);
+}
+
+.ui-request-input {
+  display: grid;
+  gap: var(--space-1);
+}
+
+.ui-request-input input {
+  min-height: var(--control-height);
+  padding: var(--space-2) var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--card);
+  color: var(--card-foreground);
+}
+
+.ui-request-actions .ui-button {
+  min-height: var(--touch-target);
+}
+
+/* Composer --------------------------------------------------------------- */
+
+.ui-composer-region {
+  display: grid;
+  gap: var(--space-2);
+}
+
+.ui-composer {
+  display: grid;
+  gap: var(--space-2);
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--card);
+}
+
+.ui-composer[data-state="active-turn"] {
+  border-color: var(--accent);
+}
+
+.ui-composer-input {
+  width: 100%;
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--card);
+  color: var(--card-foreground);
+  resize: none;
+}
+
+.ui-composer-actions {
+  display: grid;
+  gap: var(--space-2);
+}
+
+.ui-composer-hint,
+.ui-composer-reason,
+.ui-menu-reason,
+.ui-capability-diagnostic {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+}
+
+.ui-composer-reason,
+.ui-capability-diagnostic {
+  color: var(--warning);
+}
+
+.ui-composer-buttons {
+  display: flex;
+  justify-content: flex-end;
+  gap: var(--space-3);
+}
+
+.ui-composer-buttons .ui-button {
+  min-width: var(--touch-target);
+  min-height: var(--touch-target);
+}
+
+/* Banners, menus, dialogs, tooltips -------------------------------------- */
+
+.ui-banner {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  justify-content: space-between;
+  gap: var(--space-3);
+  z-index: var(--z-banner);
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--muted);
+  transition: opacity var(--motion-medium) ease;
+}
+
+.ui-banner[data-tone="accent"] {
+  border-color: var(--accent);
+  background: var(--accent-surface);
+  color: var(--accent);
+}
+
+.ui-banner[data-tone="warning"] {
+  border-color: var(--warning);
+  background: var(--warning-surface);
+  color: var(--warning);
+}
+
+.ui-banner[data-tone="danger"] {
+  border-color: var(--danger);
+  background: var(--danger-surface);
+  color: var(--danger);
+}
+
+.ui-banner-actions {
+  display: flex;
+  gap: var(--space-2);
+}
+
+.ui-menu {
+  position: relative;
+}
+
+.ui-menu-list {
+  position: absolute;
+  right: 0;
+  z-index: var(--z-banner);
+  display: grid;
+  gap: var(--space-1);
+  min-width: var(--rail-width);
+  padding: var(--space-2);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--card);
+  box-shadow: var(--shadow-sm);
+}
+
+.ui-menu-item {
+  min-height: var(--touch-target);
+  padding: var(--space-2);
+  border: 0;
+  border-radius: var(--radius-sm);
+  background: transparent;
+  color: inherit;
+  text-align: left;
+  cursor: pointer;
+}
+
+.ui-menu-item[aria-disabled="true"] {
+  color: var(--muted-foreground);
+  cursor: not-allowed;
+}
+
+.ui-menu-item:hover:not([aria-disabled="true"]) {
+  background: var(--muted);
+}
+
+.ui-dialog {
+  z-index: var(--z-dialog);
+  max-width: var(--center-width);
+  padding: 0;
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--card);
+  color: var(--card-foreground);
+}
+
+.ui-dialog-body {
+  display: grid;
+  gap: var(--space-3);
+  padding: var(--space-5);
+}
+
+.ui-dialog-title {
+  font-size: var(--font-size-lg);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-dialog-description {
+  color: var(--muted-foreground);
+}
+
+.ui-dialog-input {
+  width: 100%;
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--card);
+  color: var(--card-foreground);
+}
+
+.ui-dialog-footer {
+  display: flex;
+  justify-content: flex-end;
+  gap: var(--space-3);
+}
+
+.ui-tooltip {
+  position: relative;
+}
+
+.ui-tooltip-content {
+  position: absolute;
+  left: 0;
+  top: 100%;
+  z-index: var(--z-banner);
+  display: none;
+  min-width: var(--rail-width);
+  padding: var(--space-2);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--card);
+  font-size: var(--font-size-xs);
+}
+
+.ui-tooltip:hover .ui-tooltip-content,
+.ui-tooltip:focus-within .ui-tooltip-content {
+  display: block;
+}
+
+/* Tabs and inspector ----------------------------------------------------- */
+
+.ui-tabs {
+  display: grid;
+  gap: var(--space-3);
+}
+
+.ui-tablist {
+  display: flex;
+  flex-wrap: wrap;
+  gap: var(--space-1);
+  padding: var(--space-1);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--muted);
+}
+
+.ui-tab {
+  min-height: var(--touch-target);
+  padding-inline: var(--space-3);
+  border: 0;
+  border-radius: var(--radius-sm);
+  background: transparent;
+  color: var(--muted-foreground);
+  cursor: pointer;
+}
+
+.ui-tab[aria-selected="true"] {
+  background: var(--card);
+  color: var(--foreground);
+  font-weight: var(--weight-semibold);
+}
+
+.ui-inspector-events {
+  display: grid;
+  gap: var(--space-2);
+}
+
+.ui-inspector-filter {
+  width: 100%;
+  min-height: var(--control-height);
+  padding: var(--space-2) var(--space-3);
+  border: 0.0625rem solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--card);
+  color: var(--card-foreground);
+}
+
+.ui-facets {
+  display: flex;
+  flex-wrap: wrap;
+  gap: var(--space-1);
+  max-height: var(--facet-height);
+  overflow-y: auto;
+  padding-bottom: var(--space-1);
+}
+
+.ui-facet,
+.ui-chip,
+.ui-copy {
+  min-height: var(--space-6);
+  padding-inline: var(--space-2);
+  border: 0.0625rem solid var(--border);
+  border-radius: 999rem;
+  background: var(--card);
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+  cursor: pointer;
+}
+
+.ui-facet[aria-pressed="true"] {
+  border-color: var(--accent);
+  background: var(--accent-surface);
+  color: var(--accent);
+}
+
+.ui-inspector-list {
+  max-height: var(--transcript-height);
+  overflow-y: auto;
+}
+
+.ui-inspector-list li[data-highlighted="true"] {
+  background: var(--accent-surface);
+}
+
+.ui-inspector-row {
+  display: grid;
+  grid-template-columns: var(--timeline-column) minmax(0, 1fr) minmax(0, 1fr);
+  gap: var(--space-2);
+  align-items: center;
+  min-height: var(--touch-target);
+  cursor: pointer;
+  font-family: var(--font-mono);
+  font-size: var(--font-size-xs);
+}
+
+.ui-inspector-id {
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.ui-seq {
+  color: var(--muted-foreground);
+}
+
+.ui-inspector-grid {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr);
+  gap: var(--space-3);
+  margin: 0;
+  min-width: 0;
+}
+
+.ui-capability-list {
+  display: grid;
+  gap: var(--space-2);
+  margin: 0;
+  padding: 0;
+  list-style: none;
+}
+
+.ui-capability-list li {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr) auto;
+  gap: var(--space-2);
+  align-items: center;
+  padding-block: var(--space-2);
+  border-bottom: 0.0625rem solid var(--border);
+}
+
+.ui-capability-list p {
+  grid-column: 1 / -1;
+}
+
+.ui-replay-stepper {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: var(--space-3);
+  padding: var(--space-3);
+  border: 0.0625rem solid var(--warning);
+  border-radius: var(--radius-md);
+  background: var(--warning-surface);
+}
+
+.ui-goal-banner {
+  display: grid;
+  gap: var(--space-2);
+}
+
+.ui-goal-empty {
+  color: var(--muted-foreground);
+  font-size: var(--font-size-xs);
+}
+
+.ui-transcript li[data-highlighted="true"] {
+  outline: 0.125rem solid var(--accent);
+  outline-offset: 0.125rem;
+}
+
+.ui-tab:focus-visible,
+.ui-menu-item:focus-visible,
+.ui-tree-row:focus-visible,
+.ui-manifest-row:focus-visible,
+.ui-facet:focus-visible,
+.ui-chip:focus-visible,
+.ui-copy:focus-visible,
+.ui-jump-to-latest:focus-visible,
+.ui-request:focus-visible,
+.ui-conversation-viewport:focus-visible,
+.ui-composer-input:focus-visible,
+.ui-inspector-filter:focus-visible,
+.ui-dialog-input:focus-visible,
+.ui-request-input input:focus-visible,
+.ui-tabpanel:focus-visible,
+summary:focus-visible {
+  outline: 0.125rem solid var(--ring);
+  outline-offset: 0.125rem;
+}
+
+@media (max-width: 64rem) {
+  .ui-message[data-role="user"] {
+    margin-left: var(--space-5);
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  :root {
+    --motion-medium: 0ms;
+  }
+
+  .ui-stream-cursor {
+    animation: none;
+  }
+}
diff --git a/packages/paperclip-runner/docs/index.md b/packages/paperclip-runner/docs/index.md
index 0fdc7e17ac6..7652376b32c 100644
--- a/packages/paperclip-runner/docs/index.md
+++ b/packages/paperclip-runner/docs/index.md
@@ -11,6 +11,7 @@ Each tutorial starts with an explanation of what the phase is and what the phase
 - [Phase 3: break recovery on purpose](tutorials/phase-03-break-recovery.md) — lose an ACK, drop the socket, restart the runner side, replay durable events, and inspect recovery diagnostics.
 - [Phase 4: run the skillless Codex driver](tutorials/phase-04-skillless-codex.md) — run a safe real-model task, inspect the exact context boundary, steer or interrupt it, and confirm one replay-safe result.
 - [Phase 4b: run the protocol demo server](tutorials/phase-04b-protocol-server.md) — exercise the server-only Codex boundary, canonical replay, typed controls, and reconnect with `curl`.
+- [Phase 4b: run the live Codex protocol console](tutorials/phase-04b-live-console.md) — chat with a live session in the browser, steer it, stop it three ways, answer its requests, change its goal, break its connection, and replay the record.
 - [Cumulative end-to-end tutorial](tutorials/end-to-end.md) — the shortest complete workflow available at the current phase.
 
 ## Reference
@@ -21,11 +22,12 @@ Each tutorial starts with an explanation of what the phase is and what the phase
 - [Phase 3 durable transport and recovery](phase-03-durable-transport.md)
 - [Phase 4 skillless Codex driver](phase-04-skillless-codex-driver.md)
 - [Phase 4b protocol and demo server](phase-04b-protocol-server.md)
+- [Phase 4b live console](phase-04b-live-console.md)
 - [Engineering journal guide](journal.md)
 - [Dated shadcn/ui and AI Elements compatibility note](research/2026-08-07-ui-library-compatibility.md)
 - [Phase 4b live-console interaction map](design/phase-4b-interaction-map.md)
 - [Phase 4b component decision record (shadcn/ui, AI Elements)](design/phase-4b-component-decisions.md)
 - [Package README](../README.md)
 
-Phase 4 is implemented and ready for Security, CTO, QA, and human review.
+Phase 4b is implemented and ready for Security, CTO, QA, and human review.
 Phase 5 remains uncreated.
diff --git a/packages/paperclip-runner/docs/phase-04b-live-console.md b/packages/paperclip-runner/docs/phase-04b-live-console.md
new file mode 100644
index 00000000000..9cb3567db1a
--- /dev/null
+++ b/packages/paperclip-runner/docs/phase-04b-live-console.md
@@ -0,0 +1,138 @@
+# Phase 4b Live Console
+
+## Scope
+
+This is the browser layer of Phase 4b. It adds a **Live console** mode to the
+existing standalone devtool in `devtools/browser/`. It builds on the
+[protocol and demo server](phase-04b-protocol-server.md) and follows the
+approved [interaction map](design/phase-4b-interaction-map.md) and
+[component decision record](design/phase-4b-component-decisions.md).
+
+Every file lives under `packages/paperclip-runner/`. The console does not
+import or change Paperclip server, UI, database, or control-plane code.
+
+## Start it
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner console:phase4b
+```
+
+Then open `http://127.0.0.1:4180/` and choose **Live console**.
+
+To reach it from another host, pass a bind address to the same script:
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner console:phase4b -- --host 0.0.0.0
+```
+
+Two environment variables change the driver behind the console:
+
+| Variable | Default | Effect |
+| --- | --- | --- |
+| `PAPERCLIP_PHASE4B_DRIVER` | `demo` | `codex` starts the real `codex app-server` driver behind the same routes |
+| `PAPERCLIP_PHASE4B_CHUNK_DELAY_MS` | `45` | Milliseconds between streamed chunks in the demo driver |
+
+## Where state comes from
+
+The console has one data contract: the canonical PRP event stream and the
+`SessionSnapshot` the shared reducer produces from it. There is no second event
+model, no message store, and no client-side session cache.
+
+- The browser reduces the events it receives with the same
+  `applyPrpEvent` the server and the CLI use, so a live session and a replayed
+  session project identically. The inspector's **Session** tab reports that
+  comparison as `match` or `mismatch`.
+- Transcript order is the reducer timeline order. Nothing is re-sorted.
+- Lineage, goal state, pending requests, and the durable cursor come from the
+  demo server's public state, which is refreshed whenever new events arrive.
+- The steering chip is the only local UI intent. It resolves to
+  `acknowledged`, `rejected`, or `failed` from canonical facts and never
+  merges text into the transcript before an acknowledgement.
+
+## Surfaces
+
+| Surface | Source of truth | Notes |
+| --- | --- | --- |
+| Transcript | reducer timeline and items | user messages come from `turn.submitted` payload text |
+| Composer | composer state machine | one visible primary action: Send, Steer, or Stop |
+| Steering chip | `steering_acknowledgement` item or a rejection code | rejected text stays recoverable with **Send as new message** |
+| Request card | `runtime_request.*` events | renders only the actions the upstream request offers |
+| Goal banner and menu | capability snapshot and goal events | unsupported verbs are disabled with the exact diagnostic |
+| Lineage tree | driver thread lineage | child steering is shown as unsupported, never emulated |
+| Connection block | transport status and durable cursor | includes a deliberate connection-loss control |
+| Inspector | raw canonical events | four tabs: Events, Requests, Capabilities, Session |
+| Replay stepper | recorded canonical events | marked with a `REPLAY` eyebrow badge |
+
+## Interruption races
+
+The console distinguishes all three races the spec requires:
+
+| Race | What the console shows |
+| --- | --- |
+| interrupt before start | the pending turn resolves to `Cancelled before start`; no assistant item is ever created |
+| interrupt during generation | the stream stops in place, the partial text stays, and an interrupted divider is added |
+| interrupt during a tool call | the tool item keeps its last known status with no fabricated result |
+
+`Escape` in the composer moves focus to **Stop**. It never interrupts
+directly, because a destructive action needs an explicit activation.
+
+## Demo chats
+
+The demo server owns the manifest catalogue and serves it at
+`GET /api/phase4b/manifests`. Each manifest names the scenario it proves and
+the observations a human should confirm; the console renders that list as a
+checklist rather than restating it in the UI.
+
+| Manifest | Proves |
+| --- | --- |
+| `completion` | streaming, reasoning, tool, answer, structured result |
+| `steering` | same-turn steering acknowledgement and stale-turn rejection |
+| `interrupt-before-start` | cancellation before the provider accepts the turn |
+| `interrupt-during-generation` | preserved partial text |
+| `interrupt-during-tool` | preserved tool status with no fabricated result |
+| `approvals` | command and file-change approvals, single-response guarantee |
+| `user-input` | typed answers and request expiry |
+| `subagents` | parent/child lineage and unsupported child steering |
+| `goals` | set, view, pause, resume, clear |
+| `goals-unsupported` | exact capability diagnostic instead of a hidden control |
+| `failure` | failed item and failed turn inside the record |
+
+Manifests run in any order. **Reset demo state** names exactly what it
+discards and leaves recorded evidence files untouched.
+
+## Credentials
+
+No provider or Paperclip credential reaches the browser. The Node process owns
+the driver, the working directory, and any provider login. Every JSON and event
+frame passes the demo server's redaction layer, and the inspector renders
+redaction markers verbatim without attempting to reconstruct them.
+
+## Components
+
+Adapted from shadcn/ui and Vercel AI Elements **source**, rewritten in the
+package's local idiom, with zero new runtime dependencies:
+
+- AI Elements: `conversation`, `message`, `composer`, `reasoning-item`,
+  `tool-item`.
+- shadcn/ui: `tabs`, `menu`, `dialog`, `tooltip`, `banner`.
+
+`scripts/check-forbidden-imports.mjs` fails the build if anything under
+`devtools/browser/` imports `ai`, `@ai-sdk/*`, `zod`, `radix-ui`, `cmdk`,
+`streamdown`, `shiki`, `use-stick-to-bottom`, `class-variance-authority`,
+`tailwindcss`, or `nanoid`. `scripts/check-browser-tokens.mjs` fails if a
+component file carries a raw colour, pixel, or font value.
+
+## Verification
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner exec vitest run \
+  src/mock-core/phase4b-scripted-driver.test.ts \
+  src/mock-core/phase4b-demo-server.test.ts \
+  devtools/browser/src/live/transcript-model.test.ts
+pnpm --filter @paperclipai/paperclip-runner test:browser
+pnpm --filter @paperclipai/paperclip-runner check:browser-tokens
+pnpm --filter @paperclipai/paperclip-runner check:forbidden-imports
+```
+
+See the [verification record](../knowledge/evidence/2026-08-08-phase-04b-live-console-verification.md)
+and the [tutorial](tutorials/phase-04b-live-console.md).
diff --git a/packages/paperclip-runner/docs/tutorials/end-to-end.md b/packages/paperclip-runner/docs/tutorials/end-to-end.md
index 9b3d5ca4007..beb886f3e01 100644
--- a/packages/paperclip-runner/docs/tutorials/end-to-end.md
+++ b/packages/paperclip-runner/docs/tutorials/end-to-end.md
@@ -120,6 +120,32 @@ pnpm --filter @paperclipai/paperclip-runner verify:rootless
   browser JSON, events, and diagnostics;
 - refresh/reconnect replays canonical events and resumes the exact persisted
   provider thread.
+- the browser console renders only reducer state and canonical events, with no
+  second event model and no client-side session cache;
+- steering resolves to exactly one of acknowledged, stale-rejected, or failed,
+  and rejected text stays recoverable;
+- interrupt before start, during generation, and during a tool call each end in
+  a distinct visible state, and the session is never replaced;
+- request cards offer only the actions the upstream request offers and lock on
+  the first click until the canonical resolved event arrives;
+- unsupported capabilities render disabled controls carrying the exact upstream
+  diagnostic, never hidden and never emulated;
+- a transport drop, a page refresh, and replay all reproduce the same
+  transcript from the durable cursor;
+- no provider credential reaches the browser DOM, and adapted components add no
+  new runtime dependency.
 
-Phase 5 remains gated on the complete Phase 4b browser, review, QA, live-service,
+## Step 6: Chat with a live session in the browser
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner console:phase4b
+```
+
+Open `http://127.0.0.1:4180/` and press **Live console**. Work through the
+[Phase 4b live console tutorial](phase-04b-live-console.md) to reach every
+state above from the eleven deterministic demo chats. Add
+`PAPERCLIP_PHASE4B_DRIVER=codex` to run the identical screens against a real
+Codex session.
+
+Phase 5 remains gated on the complete Phase 4b review, QA, live-service,
 and human-checkpoint evidence.
diff --git a/packages/paperclip-runner/docs/tutorials/phase-04b-live-console.md b/packages/paperclip-runner/docs/tutorials/phase-04b-live-console.md
new file mode 100644
index 00000000000..b8b7a992758
--- /dev/null
+++ b/packages/paperclip-runner/docs/tutorials/phase-04b-live-console.md
@@ -0,0 +1,203 @@
+# Phase 4b: Run the Live Codex Protocol Console
+
+## What this is
+
+This tutorial runs the Phase 4b browser console. The
+[protocol demo server](phase-04b-protocol-server.md) tutorial used `curl`. This
+one uses the real user interface.
+
+## What this proves
+
+The console shows one live session. You can chat with it, steer it, stop it,
+answer its requests, change its goal, watch its child threads, break its
+connection, refresh the page, and replay the whole record. Every screen reads
+the canonical protocol. Nothing in the browser holds a provider login.
+
+## Before you start
+
+Run every command from the repository root. You do not need a Codex account
+for this tutorial. The default driver replays deterministic demo chats.
+
+## Step 1: Run the deterministic checks
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner exec vitest run \
+  src/mock-core/phase4b-scripted-driver.test.ts \
+  src/mock-core/phase4b-demo-server.test.ts \
+  devtools/browser/src/live/transcript-model.test.ts
+```
+
+Expected result: all three files pass. They cover the demo driver, the server
+routes, and the pure transcript model the console renders.
+
+## Step 2: Start the console
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner console:phase4b
+```
+
+Keep this terminal open. Open `http://127.0.0.1:4180/` in a browser and press
+**Live console**.
+
+Expected screen: an empty transcript that says
+`Pick a demo chat or start a blank session`, a list of demo chats on the left,
+and a disabled **Send** button.
+
+## Step 3: Run one clean turn
+
+1. Press **Completion** in the demo chat list.
+2. Press **Run this demo chat**.
+
+Expected facts:
+
+- The connection badge says `connected`.
+- Your own message appears first.
+- A collapsed **Reasoning** block appears and opens while it streams.
+- A **Tool** block shows its input and then its output.
+- The answer streams into view with a blinking cursor.
+- The turn ends with a `Turn completed` badge.
+
+## Step 4: Steer a running turn
+
+1. Press **Same-turn steering**, then **Run this demo chat**.
+2. While the turn runs, the primary button says **Steer**.
+3. Type `Keep it to one sentence.` and press Enter.
+
+Expected facts:
+
+- A steering chip appears as `pending` and then becomes `acknowledged`.
+- The chip never merges into the answer before the acknowledgement.
+
+Now type text while the turn is running, wait for `Turn completed`, and press
+the button. The chip becomes `rejected — turn already ended` and offers
+**Send as new message**. Your text is never discarded.
+
+## Step 5: Stop a turn in three ways
+
+Run each demo chat and press **Stop**:
+
+| Demo chat | When to press Stop | Expected result |
+| --- | --- | --- |
+| Interrupt before start | right away | `Cancelled before start`, and no answer item is created |
+| Interrupt during generation | while text streams | the stream stops, the partial text stays, and an interrupted divider appears |
+| Interrupt during a tool call | while the tool runs | the tool keeps its last status and shows no result |
+
+The session is never replaced. After a stop you can simply send again.
+
+## Step 6: Answer requests
+
+1. Press **Command and file approvals**, then **Run this demo chat**.
+
+Expected facts:
+
+- A blue banner above the composer says `1 request waiting — Review`.
+- The command card offers **Approve**, **Approve for session**, **Reject**,
+  and **Cancel**, because the request offers those four actions.
+- After you press **Approve**, the whole action row locks until the resolved
+  event arrives. A second click cannot answer twice.
+- The resolved card collapses to one line that names the action you chose.
+- The next card offers only **Approve** and **Reject**, because that request
+  offers only two actions.
+
+2. Press **User input and expiry**, then **Run this demo chat**.
+
+Type an answer and press **Submit**. Leave the second request alone. After a
+few seconds it becomes `expired before response` and stays in the record.
+
+## Step 7: Change the goal
+
+1. Press **Goal lifecycle**, then **Run this demo chat**.
+2. Press **Goal**, then **Set goal…**, type an objective, and confirm.
+3. Press **Goal**, then **Pause**.
+
+Expected facts: the banner shows the objective, the state, and the time of the
+last change. **Resume** is disabled while the goal is active, and **Pause** is
+disabled while it is paused.
+
+Now press **Goals unsupported** and run it. The **Goal** button is disabled,
+not hidden, and it names the exact upstream reason:
+`Goal operations unsupported: app-server 0.132.0 does not advertise the goals capability`.
+Open the **Protocol inspector** and the **Capabilities** tab. The same string
+appears there.
+
+## Step 8: Watch child threads
+
+Press **Parent and child threads**, then **Run this demo chat**.
+
+Expected facts:
+
+- The **Threads** panel lists the root session and two child threads.
+- One child ends `Completed` and one ends `Failed`. Both stay in the tree.
+- Selecting a child shows a breadcrumb and a disabled composer that says
+  `Direct steering of child threads is not supported by this app-server.`
+
+## Step 9: Break the connection and recover
+
+1. Run any demo chat and wait for it to finish.
+2. Press **Simulate connection loss**.
+
+Expected facts: a banner says the connection was lost, and the composer is
+disabled.
+
+3. Press **Retry now**.
+
+Expected facts: the badge returns to `connected` and the transcript is exactly
+what it was before the drop. Nothing is lost and nothing is duplicated.
+
+Now reload the page and press **Live console** again. The same session returns
+with its full transcript.
+
+## Step 10: Replay the record
+
+Press **Replay**.
+
+Expected facts: a `REPLAY` badge appears, the transcript resets, and a stepper
+appears. Press **Step forward** to walk the session forward one event at a
+time. Press **Play** to run it. Press **Exit replay** to return to live.
+
+## Step 11: Inspect the protocol
+
+Press **Protocol inspector**.
+
+- **Events**: every canonical event with its sequence, type, and full payload.
+  Filter by text or by event type.
+- **Requests**: every runtime request with its lifecycle.
+- **Capabilities**: what this app-server supports, with a reason for each
+  unsupported row.
+- **Session**: identities, cursors, gap state, and the live-versus-replay
+  reducer comparison. It should read `match`.
+
+Check the **Credentials in browser** row. It must read `no`.
+
+## Step 12: Check the keyboard
+
+- `Tab` reaches every control, including the inspector and the dialogs.
+- `Enter` sends. `Shift` and `Enter` add a line.
+- `Escape` while a turn runs moves focus to **Stop**. It never stops the turn
+  on its own.
+- Arrow keys move between inspector tabs and between threads.
+- `End` jumps the transcript to the newest item.
+
+## Step 13: Use the real Codex driver
+
+This step needs a working Codex login on the machine.
+
+```sh
+PAPERCLIP_PHASE4B_DRIVER=codex \
+  pnpm --filter @paperclipai/paperclip-runner console:phase4b
+```
+
+The routes and the screens are the same. The demo chat list still appears, but
+the manifest only supplies the first message; the real model produces the rest.
+The Codex login stays on the server.
+
+## Step 14: Clean up
+
+Stop the console with `Ctrl+C`. The server-owned working directory is a
+temporary directory and is not part of your repository.
+
+## Where to go next
+
+- [Cumulative end-to-end tutorial](end-to-end.md)
+- [Phase 4b live console reference](../phase-04b-live-console.md)
+- [Phase 4b interaction map](../design/phase-4b-interaction-map.md)
diff --git a/packages/paperclip-runner/knowledge/evidence/2026-08-08-phase-04b-live-console-verification.md b/packages/paperclip-runner/knowledge/evidence/2026-08-08-phase-04b-live-console-verification.md
new file mode 100644
index 00000000000..e97a55926e8
--- /dev/null
+++ b/packages/paperclip-runner/knowledge/evidence/2026-08-08-phase-04b-live-console-verification.md
@@ -0,0 +1,126 @@
+---
+type: Verification Evidence
+title: Phase 4b live console verification
+description: Deterministic driver, transcript-model, browser, boundary, and screenshot evidence for the package-local live Codex protocol console.
+tags: [native-runner, phase-4b, browser, console, evidence]
+status: stable
+generated: { by: anthropic/claude-opus-5, at: 2026-08-08T05:25:00Z }
+---
+
+# Scope
+
+This record covers the Phase 4b browser layer: the deterministic demo driver,
+the demo-chat manifests, the additive demo-server routes, and the **Live
+console** mode in `devtools/browser/`. Every file is under
+`packages/paperclip-runner/`. No Paperclip server, UI, database, or
+control-plane file is imported or changed.
+
+# Deterministic evidence
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner typecheck:typescript
+pnpm --filter @paperclipai/paperclip-runner typecheck:browser
+pnpm --filter @paperclipai/paperclip-runner exec vitest run
+```
+
+Result: typecheck clean; **14 test files, 133 tests passed**.
+
+New deterministic coverage:
+
+| File | Tests | Proves |
+| --- | --- | --- |
+| `src/mock-core/phase4b-scripted-driver.test.ts` | 15 | schema-valid canonical events, steering acknowledgement, stale-turn and already-terminal rejection, interrupt before start, preserved partial text, interrupted tool with no fabricated result, single-response requests, answers absent from acknowledgements, goal lifecycle, exact unsupported-goal diagnostic, child lineage terminal states, failed turn, resumed identity and continued source sequence |
+| `devtools/browser/src/live/transcript-model.test.ts` | 11 | role mapping, timeline-ordered transcript, submitted-message rendering, interrupted projection, resolved-request record, exact failure diagnostic, live/replay projection equality, composer state machine, goal capability gating, capability diagnostics |
+
+# Browser evidence
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner test:browser
+```
+
+Result: **28 tests passed** — 19 new live-console tests plus the 9 pre-existing
+Phase 1-3 tests, which are unchanged and still pass.
+
+The live-console tests assert, at both evidence viewports where relevant:
+
+1. empty state, streamed turn, and a completed turn with a structured result;
+2. steering acknowledgement, and a stale steer that keeps its text recoverable;
+3. all three interrupt races, each with a distinct visible outcome;
+4. approval cards that render only the offered actions, lock on first click,
+   and collapse to a named resolution;
+5. a submitted user-input answer and an unanswered request that expires;
+6. the goal lifecycle, and the exact upstream diagnostic on a disabled menu
+   mirrored in the inspector's Capabilities tab;
+7. child threads with persistent terminal states and an explicitly disabled
+   child composer carrying the exact diagnostic;
+8. a failed item with its exact diagnostic in the transcript record;
+9. a transport drop, a resume to a byte-identical transcript, and a page
+   refresh that restores the same session;
+10. replay mode with its `REPLAY` badge and stepper;
+11. the inspector, including a live-versus-replay reducer comparison reading
+    `match`;
+12. the keyboard contract: `Escape` focuses Stop without interrupting, `Enter`
+    steers, inspector tabs and the lineage tree respond to arrow keys, and the
+    transcript is a labelled `role="log"` region with `aria-live="polite"`;
+13. a reset dialog that names exactly what it discards;
+14. a 390x844 pass with no horizontal page overflow.
+
+Two negative assertions run against the rendered DOM: no `Bearer ` string and
+no `auth.json` reference reaches the browser.
+
+# Boundary and token evidence
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner check:browser-tokens
+pnpm --filter @paperclipai/paperclip-runner check:forbidden-imports
+node --test scripts/check-forbidden-imports.test.mjs
+```
+
+Result: all pass (4 boundary tests).
+
+The boundary check now rejects `ai`, `@ai-sdk/*`, `zod`, `radix-ui`,
+`@radix-ui/*`, `cmdk`, `streamdown`, `shiki`, `use-stick-to-bottom`,
+`class-variance-authority`, `tailwindcss`, and `nanoid` anywhere under
+`devtools/browser/`. A negative fixture in
+`test-fixtures/forbidden-ui-runtime/` proves the rule fires. The console adds
+**zero new runtime dependencies**.
+
+# Screenshots
+
+23 screenshots in [`phase-04b/`](phase-04b/), captured by the browser suite:
+
+- Desktop (1440x900): idle and empty, streaming turn, completed turn, steering
+  acknowledged, interrupt before start, interrupt during generation, interrupt
+  during a tool call, pending request, resolved request, expired request, goal
+  supported, goal unsupported, lineage with a child selected, failed turn,
+  reconnect banner, refresh replay, replay mode, inspector events, inspector
+  session.
+- Mobile (390x844): idle and empty, session segment, pending request,
+  inspector.
+
+# Documentation evidence
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner docs:validate
+```
+
+Result: documentation link validation and OKF v0.2 validation both pass.
+
+# Known gaps
+
+1. Real-Codex evidence for the console screens is a QA step. The scripted
+   driver proves every surface; `PAPERCLIP_PHASE4B_DRIVER=codex` runs the same
+   routes against a real session and needs a machine with a Codex login.
+2. An assistive-technology pass (screen reader) is specified in the interaction
+   map and asserted here only at the semantics level.
+3. Child-thread steering is unsupported upstream, so the console shows the
+   disabled composer rather than a live one.
+
+# Startup command for QA
+
+```sh
+pnpm --filter @paperclipai/paperclip-runner console:phase4b
+```
+
+Then open `http://127.0.0.1:4180/` and press **Live console**. Add
+`-- --host 0.0.0.0` to reach it from another host.
diff --git a/packages/paperclip-runner/knowledge/evidence/index.md b/packages/paperclip-runner/knowledge/evidence/index.md
index 1f489fff67d..4323d7d1f39 100644
--- a/packages/paperclip-runner/knowledge/evidence/index.md
+++ b/packages/paperclip-runner/knowledge/evidence/index.md
@@ -9,3 +9,4 @@
 * [Phase 3 verification](2026-08-07-phase-03-verification.md) - Lost-ACK replay, restart recovery, command deduplication, storage, and redacted diagnostics.
 * [Phase 4 verification](2026-08-08-phase-04-verification.md) - Direct Codex conformance, exact skillless context, semantic completion, recovery, and real-session replay.
 * [Phase 4b protocol/server verification](2026-08-08-phase-04b-protocol-server-verification.md) - Browser-resolved requests, goals, lineage, controls, reconnect, redaction, and real Codex server evidence.
+* [Phase 4b live console verification](2026-08-08-phase-04b-live-console-verification.md) - Demo-driver, transcript-model, browser, keyboard, boundary, token, and screenshot evidence for the live console.
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-01-idle-empty.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-01-idle-empty.png
new file mode 100644
index 00000000000..58512a34db8
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-01-idle-empty.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-02-streaming-turn.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-02-streaming-turn.png
new file mode 100644
index 00000000000..353b4bcdec4
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-02-streaming-turn.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-03-turn-completed.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-03-turn-completed.png
new file mode 100644
index 00000000000..62fe78c2120
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-03-turn-completed.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-04-steering-acknowledged.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-04-steering-acknowledged.png
new file mode 100644
index 00000000000..b2bb27e8c64
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-04-steering-acknowledged.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-05-interrupt-before-start.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-05-interrupt-before-start.png
new file mode 100644
index 00000000000..6a1e2a39257
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-05-interrupt-before-start.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-06-interrupt-during-generation.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-06-interrupt-during-generation.png
new file mode 100644
index 00000000000..5dc5d857909
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-06-interrupt-during-generation.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-07-interrupt-during-tool.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-07-interrupt-during-tool.png
new file mode 100644
index 00000000000..723c33b0a2f
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-07-interrupt-during-tool.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-08-request-pending.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-08-request-pending.png
new file mode 100644
index 00000000000..d897f663812
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-08-request-pending.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-09-request-resolved.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-09-request-resolved.png
new file mode 100644
index 00000000000..e8c91e123d3
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-09-request-resolved.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-10-request-expired.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-10-request-expired.png
new file mode 100644
index 00000000000..f058da071d7
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-10-request-expired.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-11-goal-supported.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-11-goal-supported.png
new file mode 100644
index 00000000000..8fb892eec51
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-11-goal-supported.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-12-goal-unsupported.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-12-goal-unsupported.png
new file mode 100644
index 00000000000..d65655abb93
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-12-goal-unsupported.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-13-lineage-child.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-13-lineage-child.png
new file mode 100644
index 00000000000..864be80cc64
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-13-lineage-child.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-14-failed-turn.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-14-failed-turn.png
new file mode 100644
index 00000000000..08c66b7f1a8
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-14-failed-turn.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-15-reconnect-banner.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-15-reconnect-banner.png
new file mode 100644
index 00000000000..342d25dfc0f
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-15-reconnect-banner.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-16-refresh-replay.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-16-refresh-replay.png
new file mode 100644
index 00000000000..41479c51b66
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-16-refresh-replay.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-17-replay-mode.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-17-replay-mode.png
new file mode 100644
index 00000000000..af7f4ebd9c0
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-17-replay-mode.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-18-inspector-events.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-18-inspector-events.png
new file mode 100644
index 00000000000..2e301c76cca
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-18-inspector-events.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-19-inspector-session.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-19-inspector-session.png
new file mode 100644
index 00000000000..a40750be631
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/desktop-19-inspector-session.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-01-idle-empty.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-01-idle-empty.png
new file mode 100644
index 00000000000..57012d2bda0
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-01-idle-empty.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-02-session-segment.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-02-session-segment.png
new file mode 100644
index 00000000000..bce20593ef2
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-02-session-segment.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-03-request-pending.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-03-request-pending.png
new file mode 100644
index 00000000000..920fb07b4d0
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-03-request-pending.png differ
diff --git a/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-04-inspector.png b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-04-inspector.png
new file mode 100644
index 00000000000..bbd82983413
Binary files /dev/null and b/packages/paperclip-runner/knowledge/evidence/phase-04b/mobile-04-inspector.png differ
diff --git a/packages/paperclip-runner/knowledge/journal/2026-08-08-phase-04b-live-console.md b/packages/paperclip-runner/knowledge/journal/2026-08-08-phase-04b-live-console.md
new file mode 100644
index 00000000000..ac3e243375f
--- /dev/null
+++ b/packages/paperclip-runner/knowledge/journal/2026-08-08-phase-04b-live-console.md
@@ -0,0 +1,96 @@
+---
+type: Engineering Journal Entry
+title: Phase 4b live console browser tracer
+description: Decisions, failures, fixes, and evidence for the package-local browser console over the canonical runner protocol.
+tags: [native-runner, phase-4b, browser, console, reducer]
+status: stable
+generated: { by: anthropic/claude-opus-5, at: 2026-08-08T05:25:00Z }
+entry_kind: phase
+phase: "4b"
+---
+
+# Context
+
+The [protocol and demo-server layer](2026-08-08-phase-04b-protocol-server.md)
+gave the browser typed routes and a server-only credential boundary. The
+[UX gate](2026-08-08-phase-04b-ux.md) gave it an approved interaction map and a
+component decision record. This entry records building the console itself.
+
+# Decisions
+
+1. Extend the existing devtool with a **Live console** mode. Do not build a
+   second app shell.
+2. Keep one data contract: canonical PRP events plus the shared reducer. The
+   browser runs `applyPrpEvent`, the same function the server and CLI run, so
+   live output and replayed output cannot drift.
+3. Add a deterministic scripted `HarnessDriver` behind the same demo-server
+   routes the real Codex driver uses. The console can then reach every
+   protocol state without a provider account, and the real driver is a single
+   environment variable away.
+4. Let the manifest, not the UI, own the list of expected observations. The
+   human checkpoint script lives in the data, not in hand-written copy.
+5. Adapt shadcn/ui and AI Elements source with zero new runtime dependencies.
+   Rebuild the radix-backed patterns on native primitives: `<dialog>` for
+   modals, `<details>` for disclosures, the WAI-ARIA tabs and menu-button
+   patterns by hand.
+6. Put the submitted message text into the `turn.submitted` payload so the
+   transcript can show the operator's own message from a canonical event
+   instead of a shadow store.
+7. Render exactly one layout. A second CSS-hidden copy would duplicate the
+   transcript log landmark and every control in the DOM.
+
+# Failures and fixes
+
+1. **The mobile and desktop trees both rendered.** The first build hid one with
+   a media query. That duplicated every `data-testid`, every interactive
+   control, and the `role="log"` landmark. Fixed with a `matchMedia` hook that
+   renders one tree.
+2. **A resumed stream never opened.** The SSE route wrote its headers and then
+   waited. When a subscriber resumed past the last event there were no bytes to
+   send, so Node held the response open and the browser stayed in `CONNECTING`
+   forever. Fixed by writing an SSE comment immediately after the headers.
+   This only appeared after a reconnect, which is exactly the case the phase is
+   about.
+3. **A finished turn stayed steerable.** The composer read
+   `snapshot.activeTurnId ?? state.activeTurnId`. When the reducer correctly
+   set the active turn to `null`, the nullish fallback reached back to a stale
+   public-state read and the primary action stayed **Steer**. Fixed by making
+   the reducer authoritative whenever a snapshot exists.
+4. **Interrupt before start was unreachable.** The composer left `submitting`
+   as soon as the create call returned, so **Stop** was disabled during the
+   exact window the race needs. Fixed by deriving the pending-turn state from
+   the timeline: a `turn.submitted` with no later `turn.*` event.
+5. **Lineage, goals, and pending requests went stale.** The event stream
+   carries events; those fields live in the server's public state. Fixed with a
+   debounced state refresh whenever new events arrive.
+6. **The left rail overflowed its track.** Long run and session identifiers
+   gave the rail panels a min-content width of 398px inside a 272px track, and
+   the panels slid under the transcript. Fixed with `minmax(0, 1fr)` tracks and
+   wrapping identifiers. Found by measuring bounding boxes in the browser, not
+   by reading the CSS.
+7. **The new mode broke five frozen Phase 1-3 browser tests.** Landing on the
+   console changed the entry path those specs rely on. Reverted the landing
+   mode and navigated explicitly in the new spec instead.
+
+# Evidence
+
+- 15 driver and manifest tests, 11 transcript-model tests, 19 live-console
+  browser tests, and the 9 pre-existing browser tests all pass.
+- 23 screenshots at 1440x900 and 390x844 in
+  [`evidence/phase-04b/`](../evidence/phase-04b/).
+- The boundary check now fails on any AI SDK, radix, Tailwind, or markdown
+  runtime import under `devtools/browser/`, with a negative fixture proving it.
+- Full record: [Phase 4b live console verification](../evidence/2026-08-08-phase-04b-live-console-verification.md).
+
+# Gaps and next questions
+
+1. Child threads are read-only because the app-server does not advertise child
+   steering. If a later Codex release adds it, the disabled composer becomes a
+   capability-gated live one; the diagnostic string is already the seam.
+2. Real-Codex evidence for the console UI is a QA step, not an automated one.
+   The scripted driver proves the surfaces; a human still has to confirm the
+   same screens against a real session.
+3. Screen-reader passes are specified in the interaction map and asserted at
+   the semantics level here. An assistive-technology pass remains a QA task.
+4. Markdown rendering and syntax highlighting stay rejected. If Phase 5 turns
+   this into an SDK console, both become dependency decisions to revisit.
diff --git a/packages/paperclip-runner/knowledge/journal/index.md b/packages/paperclip-runner/knowledge/journal/index.md
index 50070150f00..bd4b9aca4d5 100644
--- a/packages/paperclip-runner/knowledge/journal/index.md
+++ b/packages/paperclip-runner/knowledge/journal/index.md
@@ -10,3 +10,4 @@
 * [Phase 4 skillless Codex app-server driver](2026-08-08-phase-04.md) - Direct app-server mapping, context isolation, semantic completion, real-model failures, and fixes.
 * [Phase 4b UX gate — interaction map and component decisions](2026-08-08-phase-04b-ux.md) - Approved live-console interaction contract, adapt/reject component record, and keyboard/accessibility acceptance criteria.
 * [Phase 4b protocol and demo-server layer](2026-08-08-phase-04b-protocol-server.md) - Browser-resolved provider requests, control races, goals, lineage, replay, and server-only credentials.
+* [Phase 4b live console browser tracer](2026-08-08-phase-04b-live-console.md) - Reducer-only browser surfaces, adapted components with no new runtime dependencies, interrupt races, reconnect, replay, and the layout and stream failures found on the way.
diff --git a/packages/paperclip-runner/knowledge/log.md b/packages/paperclip-runner/knowledge/log.md
index 87bf9df0845..0448f5d3ea2 100644
--- a/packages/paperclip-runner/knowledge/log.md
+++ b/packages/paperclip-runner/knowledge/log.md
@@ -21,3 +21,11 @@
 * **Phase 4b protocol/server layer**: Added deterministic Codex request, goal,
   lineage, control-race, reconnect, replay, and redaction fixtures; a typed
   server-only demo boundary; and a passing real Codex server trace.
+* **Phase 4b live console**: Added the package-local browser tracer over the
+  canonical protocol — transcript, composer, same-turn steering, three
+  interrupt races, runtime request cards, capability-gated goals, parent/child
+  lineage, reconnect, refresh, replay, and a redacted protocol inspector — plus
+  a deterministic demo driver, eleven demo chats, adapted shadcn/AI Elements
+  components with no new runtime dependencies, and a boundary check that keeps
+  it that way.
+
diff --git a/packages/paperclip-runner/package.json b/packages/paperclip-runner/package.json
index 3aec9c9d11c..0f9955c0599 100644
--- a/packages/paperclip-runner/package.json
+++ b/packages/paperclip-runner/package.json
@@ -52,6 +52,7 @@
     "diagnose:phase3": "pnpm run trace:phase3",
     "record:phase3": "pnpm run build:typescript && pnpm run build:phase2-binaries && node scripts/record-phase3-evidence.mjs",
     "record:phase3:lost-ack": "pnpm run trace:phase3 -- --fault lost-ack --output knowledge/evidence/phase-03-lost-ack.json",
+    "console:phase4b": "pnpm run build:typescript && vite --config vite.config.ts --host 127.0.0.1 --port 4180",
     "browser:dev": "pnpm run build:typescript && pnpm run build:phase2-binaries && vite --config vite.config.ts",
     "browser:preview": "vite preview --config vite.config.ts",
     "verify": "pnpm run build && pnpm run typecheck && pnpm run check:phase1-goldens && pnpm run test && pnpm run check:browser-tokens && pnpm run test:browser && pnpm run check:forbidden-imports && pnpm run docs:validate && pnpm run check:phase0-parity && pnpm run check:phase1-parity && pnpm run trace:phase0 && pnpm run replay:phase1 && pnpm run trace:phase2 -- --quiet && pnpm run trace:phase3",
diff --git a/packages/paperclip-runner/scripts/check-forbidden-imports.test.mjs b/packages/paperclip-runner/scripts/check-forbidden-imports.test.mjs
index 3a6035f7f68..6c48f5e81e8 100644
--- a/packages/paperclip-runner/scripts/check-forbidden-imports.test.mjs
+++ b/packages/paperclip-runner/scripts/check-forbidden-imports.test.mjs
@@ -35,3 +35,16 @@ test("a negative Cargo fixture proves that a core path dependency is rejected",
   assert.match(violations[0].reason, /Cargo path dependencies may not escape/);
   assert.ok(violations[0].file.startsWith(defaultPackageRoot));
 });
+
+test("a negative fixture proves that a browser UI runtime import is rejected", async () => {
+  const violations = await checkForbiddenImports({
+    scanRoots: ["devtools/browser", "test-fixtures/forbidden-ui-runtime/devtools/browser"],
+    cargoRoots: [],
+    checkManifest: false,
+  });
+
+  assert.equal(violations.length, 1);
+  assert.equal(violations[0].specifier, "@ai-sdk/react");
+  assert.match(violations[0].reason, /adapts component source/);
+  assert.ok(violations[0].file.startsWith(defaultPackageRoot));
+});
diff --git a/packages/paperclip-runner/scripts/lib/forbidden-imports.mjs b/packages/paperclip-runner/scripts/lib/forbidden-imports.mjs
index 422b615d1d6..0c67c1d2dfd 100644
--- a/packages/paperclip-runner/scripts/lib/forbidden-imports.mjs
+++ b/packages/paperclip-runner/scripts/lib/forbidden-imports.mjs
@@ -55,11 +55,42 @@ function findSpecifiers(source) {
   return found;
 }
 
+// The Phase 4b component decision record adapts shadcn/ui and AI Elements
+// source rather than adopting their runtimes. These packages would reintroduce
+// a second message model or a Tailwind/radix dependency for the demo app.
+const BROWSER_FORBIDDEN_PACKAGES = [
+  "ai",
+  "@ai-sdk",
+  "zod",
+  "radix-ui",
+  "@radix-ui",
+  "cmdk",
+  "streamdown",
+  "shiki",
+  "use-stick-to-bottom",
+  "class-variance-authority",
+  "tailwindcss",
+  "nanoid",
+];
+
+function isForbiddenBrowserPackage(specifier) {
+  return BROWSER_FORBIDDEN_PACKAGES.some(
+    (name) => specifier === name || specifier.startsWith(`${name}/`),
+  );
+}
+
 function violationReason({ file, packageRoot, specifier }) {
   if (specifier.startsWith("@paperclipai/") && specifier !== "@paperclipai/paperclip-runner") {
     return "Paperclip workspace packages are outside the standalone boundary";
   }
 
+  if (
+    relative(packageRoot, file).split(/[\\/]/).join("/").includes("devtools/browser/") &&
+    isForbiddenBrowserPackage(specifier)
+  ) {
+    return "the standalone browser app adapts component source instead of adopting its runtime";
+  }
+
   if (["server", "ui", "cli"].some((root) => specifier === root || specifier.startsWith(`${root}/`))) {
     return "Paperclip application internals are outside the standalone boundary";
   }
diff --git a/packages/paperclip-runner/scripts/phase4b-browser-server.mjs b/packages/paperclip-runner/scripts/phase4b-browser-server.mjs
new file mode 100644
index 00000000000..b0f3e564a46
--- /dev/null
+++ b/packages/paperclip-runner/scripts/phase4b-browser-server.mjs
@@ -0,0 +1,140 @@
+import { mkdtemp } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { resolve } from "node:path";
+
+/**
+ * Mounts the Phase 4b demo-server routes inside the standalone Vite devtool.
+ *
+ * The browser never talks to a provider. Only this Node process starts a
+ * driver, owns the working directory, and reads any provider login. The
+ * default driver replays deterministic demo manifests; setting
+ * `PAPERCLIP_PHASE4B_DRIVER=codex` swaps in the real Codex app-server driver
+ * behind exactly the same routes.
+ */
+const LOOPBACK_ONLY = "Phase 4b is available only over loopback";
+
+function isLoopbackAddress(address) {
+  if (typeof address !== "string") return false;
+  const normalized = address.toLowerCase().replace(/^::ffff:/, "");
+  if (normalized === "::1" || normalized === "[::1]" || normalized === "localhost") return true;
+  const octets = normalized.split(".");
+  return (
+    octets.length === 4 &&
+    octets[0] === "127" &&
+    octets.every((octet) => /^\d{1,3}$/.test(octet) && Number(octet) <= 255)
+  );
+}
+
+function rejectUntrusted(request, response) {
+  if (
+    !isLoopbackAddress(request.socket.localAddress) ||
+    !isLoopbackAddress(request.socket.remoteAddress)
+  ) {
+    response.statusCode = 403;
+    response.setHeader("content-type", "application/json; charset=utf-8");
+    response.end(JSON.stringify({ error: "forbidden", message: LOOPBACK_ONLY }));
+    return true;
+  }
+  const fetchSite = request.headers["sec-fetch-site"];
+  if (
+    Array.isArray(fetchSite) ||
+    (typeof fetchSite === "string" && !["same-origin", "none"].includes(fetchSite))
+  ) {
+    response.statusCode = 403;
+    response.setHeader("content-type", "application/json; charset=utf-8");
+    response.end(JSON.stringify({
+      error: "forbidden",
+      message: "Phase 4b cross-site requests are forbidden",
+    }));
+    return true;
+  }
+  return false;
+}
+
+async function loadRunner() {
+  return import(new URL("../dist/index.js", import.meta.url).href);
+}
+
+async function createWorkingDirectory() {
+  const scratchRoot =
+    process.env.PAPERCLIP_RUN_SCRATCH_DIR ?? process.env.PAPERCLIP_SCRATCH_DIR ?? tmpdir();
+  return mkdtemp(resolve(scratchRoot, "phase4b-browser-"));
+}
+
+export function createPhase4bBrowserMiddleware(options = {}) {
+  const load = options.loadRunner ?? loadRunner;
+  const driverMode = options.driverMode ?? process.env.PAPERCLIP_PHASE4B_DRIVER ?? "demo";
+  const chunkDelayMs = Number.parseInt(
+    options.chunkDelayMs ?? process.env.PAPERCLIP_PHASE4B_CHUNK_DELAY_MS ?? "45",
+    10,
+  );
+  let bootstrap = null;
+
+  async function ready() {
+    if (bootstrap !== null) return bootstrap;
+    bootstrap = (async () => {
+      const runner = await load();
+      const workingDirectory =
+        options.workingDirectory ?? (await createWorkingDirectory());
+      const server = new runner.Phase4bDemoServer({
+        workingDirectory,
+        manifests: runner.phase4bDemoManifestCatalogue(),
+        driverFactory: (taskEnvelope, manifestId) =>
+          driverMode === "codex"
+            ? new runner.CodexAppServerDriver({
+                taskEnvelope,
+                onDiagnostic: (message) => process.stderr.write(`[phase4b] ${message}\n`),
+              })
+            : new runner.Phase4bScriptedDriver({
+                manifestId: manifestId ?? "completion",
+                chunkDelayMs: Number.isInteger(chunkDelayMs) ? chunkDelayMs : 45,
+              }),
+      });
+      return { handle: server.middleware(), server, workingDirectory };
+    })();
+    return bootstrap;
+  }
+
+  const middleware = async function phase4bMiddleware(request, response, next) {
+    const url = new URL(request.url ?? "/", "http://phase4b.local");
+    if (!url.pathname.startsWith("/api/phase4b/")) {
+      next();
+      return;
+    }
+    if (rejectUntrusted(request, response)) return;
+    try {
+      const { handle } = await ready();
+      handle(request, response);
+    } catch (error) {
+      response.statusCode = 500;
+      response.setHeader("content-type", "application/json; charset=utf-8");
+      response.end(JSON.stringify({
+        error: "phase4b_unavailable",
+        message: String(error instanceof Error ? error.message : error),
+      }));
+    }
+  };
+  middleware.close = async () => {
+    if (bootstrap === null) return;
+    const { server } = await bootstrap;
+    await server.close();
+  };
+  return middleware;
+}
+
+export function phase4bBrowserServerPlugin(options = {}) {
+  const middleware = createPhase4bBrowserMiddleware(options);
+  return {
+    name: "paperclip-runner-phase4b-live-console-server",
+    configureServer(server) {
+      server.middlewares.use(middleware);
+      server.httpServer?.once("close", () => void middleware.close());
+    },
+    configurePreviewServer(server) {
+      server.middlewares.use(middleware);
+      server.httpServer?.once("close", () => void middleware.close());
+    },
+  };
+}
+
+export const phase4bBrowserServerInternals = { isLoopbackAddress };
diff --git a/packages/paperclip-runner/src/drivers/codex/codex-app-server-driver.ts b/packages/paperclip-runner/src/drivers/codex/codex-app-server-driver.ts
index b7c6219808c..0244d539ca2 100644
--- a/packages/paperclip-runner/src/drivers/codex/codex-app-server-driver.ts
+++ b/packages/paperclip-runner/src/drivers/codex/codex-app-server-driver.ts
@@ -709,7 +709,13 @@ class CodexHarnessSession implements HarnessSession {
       );
     }
     const taskText = JSON.stringify({ task: this.#taskEnvelope, message: input.message.text });
-    this.#emit("turn.submitted", { envelopeSchema: this.#taskEnvelope.schema });
+    // The submitted text is part of the canonical record so a tracer can show
+    // the operator's own message without keeping shadow state next to the
+    // reducer.
+    this.#emit("turn.submitted", {
+      envelopeSchema: this.#taskEnvelope.schema,
+      text: input.message.text,
+    });
     this.#turnStartPending = true;
     let response: Record<string, unknown>;
     try {
diff --git a/packages/paperclip-runner/src/index.ts b/packages/paperclip-runner/src/index.ts
index 9fa8c658d2a..205ea23e975 100644
--- a/packages/paperclip-runner/src/index.ts
+++ b/packages/paperclip-runner/src/index.ts
@@ -10,6 +10,8 @@ export * from "./drivers/codex/codex-app-server-driver.js";
 export * from "./mock-core/mock-control-plane-adapter.js";
 export * from "./mock-core/phase4-codex-runner.js";
 export * from "./mock-core/phase4b-demo-server.js";
+export * from "./mock-core/phase4b-demo-manifests.js";
+export * from "./mock-core/phase4b-scripted-driver.js";
 export * from "./mock-core/phase2-local-runner.js";
 export * from "./mock-core/phase3-recovery.js";
 export * from "./protocol/phase0-fixture.js";
diff --git a/packages/paperclip-runner/src/mock-core/phase4b-demo-manifests.ts b/packages/paperclip-runner/src/mock-core/phase4b-demo-manifests.ts
new file mode 100644
index 00000000000..817e9201bc5
--- /dev/null
+++ b/packages/paperclip-runner/src/mock-core/phase4b-demo-manifests.ts
@@ -0,0 +1,448 @@
+import type { HarnessRuntimeRequestKind } from "../contracts/harness-driver.js";
+import type { NativeSessionCapabilities } from "../contracts/types.js";
+
+/**
+ * A demo chat manifest is a deterministic script for the Phase 4b browser
+ * tracer. Every state named in the interaction map has to be reachable from
+ * one of these manifests, and the manifest — not the browser — owns the list
+ * of observations a human is expected to confirm.
+ */
+export type Phase4bDemoBeat =
+  | { kind: "reasoning"; itemId: string; chunks: readonly string[] }
+  | { kind: "assistant"; itemId: string; chunks: readonly string[] }
+  | {
+      kind: "tool";
+      itemId: string;
+      name: string;
+      input: string;
+      output?: string;
+      failure?: string;
+      /** Stay in flight until an interrupt or the bounded demo timeout. */
+      hold?: boolean;
+    }
+  | {
+      kind: "request";
+      requestId: string;
+      requestKind: HarnessRuntimeRequestKind;
+      method: string;
+      itemId: string;
+      prompt: string;
+      details: Record<string, unknown>;
+      /** Resolution actions the upstream request actually offers. */
+      actions: readonly string[];
+      /** Expire the request when nobody answers within the bounded window. */
+      expiresAfterMs?: number;
+    }
+  | {
+      kind: "child";
+      threadId: string;
+      nickname: string;
+      role: string;
+      statuses: readonly string[];
+    }
+  | { kind: "hold"; note: string }
+  | { kind: "diagnostic"; message: string };
+
+export interface Phase4bDemoManifest {
+  id: string;
+  name: string;
+  purpose: string;
+  scenarios: readonly string[];
+  objective: string;
+  prompt: string;
+  capabilities: NativeSessionCapabilities;
+  /** Milliseconds between `turn.submitted` and `turn.accepted`. */
+  startDelayMs: number;
+  beats: readonly Phase4bDemoBeat[];
+  /** Beats used by every turn after the first one. */
+  followUpBeats?: readonly Phase4bDemoBeat[];
+  expectedObservations: readonly string[];
+  disposition: "completed" | "failed";
+}
+
+const FULL_CAPABILITIES: NativeSessionCapabilities = {
+  resume: true,
+  typedEvents: true,
+  steering: true,
+  interruption: true,
+  structuredResult: true,
+  read: true,
+  reconciliation: true,
+  usage: true,
+  runtimeRequestResolution: true,
+  goals: true,
+  threadLineage: true,
+};
+
+function capabilities(
+  overrides: Partial<NativeSessionCapabilities> = {},
+): NativeSessionCapabilities {
+  return { ...FULL_CAPABILITIES, ...overrides };
+}
+
+const SHORT_ANSWER: readonly string[] = [
+  "I read the objective, ",
+  "checked the working directory, ",
+  "and finished the task.",
+];
+
+export const PHASE4B_DEMO_MANIFESTS: readonly Phase4bDemoManifest[] = [
+  {
+    id: "completion",
+    name: "Completion",
+    purpose: "One clean turn: reasoning, a tool call, an answer, and a terminal result.",
+    scenarios: ["transcript", "streaming", "terminal"],
+    objective: "Summarize the Phase 4b boundary.",
+    prompt: "Summarize what the Phase 4b demo server proves.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "reasoning",
+        itemId: "reasoning-1",
+        chunks: ["Reading the objective. ", "The task is a summary, so no writes are needed."],
+      },
+      {
+        kind: "tool",
+        itemId: "tool-1",
+        name: "read_file",
+        input: "docs/phase-04b-protocol-server.md",
+        output: "The browser boundary is HTTP plus server-sent events.",
+      },
+      { kind: "assistant", itemId: "assistant-1", chunks: SHORT_ANSWER },
+    ],
+    followUpBeats: [
+      { kind: "assistant", itemId: "assistant-follow", chunks: ["Follow-up turns reuse the same session."] },
+    ],
+    expectedObservations: [
+      "The transcript shows reasoning collapsed and the answer expanded.",
+      "The tool item shows its input and output in a monospace block.",
+      "The turn ends with a completed badge and a structured result.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "steering",
+    name: "Same-turn steering",
+    purpose: "Steer a running turn, watch the acknowledgement, then watch a stale steer get rejected.",
+    scenarios: ["steering", "stale-turn"],
+    objective: "Keep the answer short.",
+    prompt: "Explain the steering contract, then wait for further direction.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "reasoning",
+        itemId: "reasoning-1",
+        chunks: ["Starting a long answer so there is time to steer."],
+      },
+      {
+        kind: "assistant",
+        itemId: "assistant-1",
+        chunks: [
+          "Steering is bound to the turn that was active when the message was composed. ",
+          "The driver compares the expected turn id before it accepts the text. ",
+          "A mismatch is rejected instead of being applied to a different turn.",
+        ],
+      },
+    ],
+    expectedObservations: [
+      "A steering chip appears as pending and then becomes acknowledged.",
+      "Steering sent after the turn ends is rejected as a stale turn.",
+      "The rejected chip keeps the typed text and offers to send it as a new message.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "interrupt-before-start",
+    name: "Interrupt before start",
+    purpose: "Press stop while the turn is still waiting for a provider turn id.",
+    scenarios: ["interrupt"],
+    objective: "Cancel before the provider starts.",
+    prompt: "Start a slow turn so it can be cancelled before it starts.",
+    capabilities: capabilities(),
+    startDelayMs: 4_000,
+    beats: [{ kind: "assistant", itemId: "assistant-1", chunks: ["The turn started after all."] }],
+    expectedObservations: [
+      "The composer sits in the submitting state while the turn waits to start.",
+      "Stop resolves the pending turn to `Cancelled before start`.",
+      "No assistant item is ever created for the cancelled turn.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "interrupt-during-generation",
+    name: "Interrupt during generation",
+    purpose: "Stop a streaming answer and keep the partial text.",
+    scenarios: ["interrupt", "streaming"],
+    objective: "Stop mid-answer.",
+    prompt: "Write a long answer so it can be interrupted while streaming.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "assistant",
+        itemId: "assistant-1",
+        chunks: [
+          "The tracer keeps every partial token that arrived before the stop. ",
+          "It never blanks an interrupted item, ",
+          "because the partial record is the point of a tracer. ",
+          "This sentence exists so the stream lasts long enough to interrupt. ",
+          "And this one too. ",
+          "And this final one.",
+        ],
+      },
+    ],
+    expectedObservations: [
+      "The stream stops in place and the partial text stays visible.",
+      "The item shows an interrupted divider instead of a completed state.",
+      "The turn group shows an interrupted badge.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "interrupt-during-tool",
+    name: "Interrupt during a tool call",
+    purpose: "Stop while a tool call is running and keep its last known status.",
+    scenarios: ["interrupt", "tools"],
+    objective: "Stop during a tool call.",
+    prompt: "Run a slow tool so it can be interrupted.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "tool",
+        itemId: "tool-1",
+        name: "run_command",
+        input: "sleep 30 && echo done",
+        hold: true,
+      },
+      { kind: "assistant", itemId: "assistant-1", chunks: ["The tool finished."] },
+    ],
+    expectedObservations: [
+      "The tool item stays running until the stop is pressed.",
+      "The interrupted tool item keeps its last known status and no fabricated result.",
+      "The turn group shows an interrupted badge.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "approvals",
+    name: "Command and file approvals",
+    purpose: "Resolve a command approval and a file-change approval from the transcript.",
+    scenarios: ["requests", "approvals"],
+    objective: "Approve the demo command and file change.",
+    prompt: "Ask for approval before running the command and writing the file.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "request",
+        requestId: "request-command",
+        requestKind: "command_approval",
+        method: "item/commandExecution/requestApproval",
+        itemId: "command-1",
+        prompt: "Run the deterministic demo command?",
+        details: { command: "node scripts/phase4b-demo-check.mjs --dry-run", cwd: "<server-path>" },
+        actions: ["accept", "accept_for_session", "decline", "cancel"],
+      },
+      {
+        kind: "request",
+        requestId: "request-file",
+        requestKind: "file_approval",
+        method: "item/fileChange/requestApproval",
+        itemId: "file-1",
+        prompt: "Write phase4b-demo.txt?",
+        details: {
+          path: "phase4b-demo.txt",
+          diff: "+phase4b demo evidence\n",
+          changeKind: "add",
+        },
+        actions: ["accept", "decline"],
+      },
+      { kind: "assistant", itemId: "assistant-1", chunks: ["Both approvals were honoured exactly once."] },
+    ],
+    expectedObservations: [
+      "Each request card renders only the actions the request offers.",
+      "The first click disables the whole action row until the resolved event arrives.",
+      "A resolved card collapses to one summary line that names the chosen action.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "user-input",
+    name: "User input and expiry",
+    purpose: "Answer a typed user-input request, then watch an unanswered request expire.",
+    scenarios: ["requests", "user-input", "expiry"],
+    objective: "Collect the deployment target, then let a request expire.",
+    prompt: "Ask which environment to use, then ask a question nobody answers.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "request",
+        requestId: "request-input",
+        requestKind: "user_input",
+        method: "item/tool/requestUserInput",
+        itemId: "input-1",
+        prompt: "Which environment should the demo target?",
+        details: { fields: [{ name: "environment", label: "Environment" }] },
+        actions: ["submit", "decline", "cancel"],
+      },
+      {
+        kind: "request",
+        requestId: "request-expiring",
+        requestKind: "elicitation",
+        method: "mcpServer/elicitation/request",
+        itemId: "input-2",
+        prompt: "Optional detail — this request expires if nobody answers.",
+        details: { fields: [{ name: "detail", label: "Detail" }] },
+        actions: ["submit", "decline", "cancel"],
+        expiresAfterMs: 8_000,
+      },
+      { kind: "assistant", itemId: "assistant-1", chunks: ["The expired request stays in the record."] },
+    ],
+    expectedObservations: [
+      "The user-input card renders a text field and a submit action.",
+      "The submitted answer never appears in the resolution acknowledgement.",
+      "The unanswered request resolves to `expired before response` on its own.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "subagents",
+    name: "Parent and child threads",
+    purpose: "Watch child threads start, report, and reach a terminal state.",
+    scenarios: ["lineage", "subagents"],
+    objective: "Delegate two checks to child threads.",
+    prompt: "Spawn two child threads and report what they found.",
+    capabilities: capabilities({ steering: true }),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "child",
+        threadId: "thread-child-a",
+        nickname: "docs-check",
+        role: "reviewer",
+        statuses: ["starting", "running", "completed"],
+      },
+      {
+        kind: "child",
+        threadId: "thread-child-b",
+        nickname: "test-check",
+        role: "reviewer",
+        statuses: ["starting", "running", "failed"],
+      },
+      { kind: "assistant", itemId: "assistant-1", chunks: ["One child finished and one child failed."] },
+    ],
+    expectedObservations: [
+      "The lineage tree lists the root session and both child threads.",
+      "Child terminal states stay in the tree after the turn ends.",
+      "Selecting a child shows an explicitly disabled composer with the exact diagnostic.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "goals",
+    name: "Goal lifecycle",
+    purpose: "Set, pause, resume, and clear a durable goal.",
+    scenarios: ["goals", "capabilities"],
+    objective: "Keep the demo goal visible.",
+    prompt: "Work under a durable goal.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      { kind: "assistant", itemId: "assistant-1", chunks: ["Use the goal menu to change the goal state."] },
+    ],
+    expectedObservations: [
+      "The goal banner shows the active goal and its last transition time.",
+      "Pause and resume are disabled when they are invalid for the current state.",
+      "Every goal change is also a transcript event, so replay shows the same history.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "goals-unsupported",
+    name: "Goals unsupported",
+    purpose: "Show the exact capability diagnostic instead of hiding the control.",
+    scenarios: ["goals", "capabilities"],
+    objective: "Prove unsupported capabilities are visible.",
+    prompt: "Run against an app-server that does not advertise goals.",
+    capabilities: capabilities({
+      goals: false,
+      unsupported: [
+        "Goal operations unsupported: app-server 0.132.0 does not advertise the goals capability",
+      ],
+    }),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "assistant",
+        itemId: "assistant-1",
+        chunks: ["The goal menu stays visible and disabled with the upstream diagnostic."],
+      },
+    ],
+    expectedObservations: [
+      "The goal menu is disabled, not hidden.",
+      "The exact capability diagnostic is readable without hovering.",
+      "The inspector capabilities tab shows the same unsupported string.",
+    ],
+    disposition: "completed",
+  },
+  {
+    id: "failure",
+    name: "Item and turn failure",
+    purpose: "Show a failed item and a failed turn as part of the record.",
+    scenarios: ["failure", "diagnostics"],
+    objective: "Fail on purpose.",
+    prompt: "Run a command that fails.",
+    capabilities: capabilities(),
+    startDelayMs: 0,
+    beats: [
+      {
+        kind: "tool",
+        itemId: "tool-1",
+        name: "run_command",
+        input: "node scripts/missing.mjs",
+        failure: "Command exited with status 1: cannot find module scripts/missing.mjs",
+      },
+      { kind: "diagnostic", message: "The demo task cannot continue after the failed command." },
+    ],
+    expectedObservations: [
+      "The failed item renders in place with the exact diagnostic string.",
+      "No failure is reported only as a transient notification.",
+      "The turn ends in a failed state and the run result reports the failure.",
+    ],
+    disposition: "failed",
+  },
+];
+
+export function findPhase4bDemoManifest(id: string): Phase4bDemoManifest {
+  const manifest = PHASE4B_DEMO_MANIFESTS.find((candidate) => candidate.id === id);
+  if (manifest === undefined) {
+    throw new Error(`unknown Phase 4b demo manifest ${id}`);
+  }
+  return manifest;
+}
+
+export interface Phase4bDemoManifestSummary {
+  id: string;
+  name: string;
+  purpose: string;
+  scenarios: readonly string[];
+  objective: string;
+  prompt: string;
+  expectedObservations: readonly string[];
+}
+
+export function phase4bDemoManifestCatalogue(): Phase4bDemoManifestSummary[] {
+  return PHASE4B_DEMO_MANIFESTS.map((manifest) => ({
+    id: manifest.id,
+    name: manifest.name,
+    purpose: manifest.purpose,
+    scenarios: [...manifest.scenarios],
+    objective: manifest.objective,
+    prompt: manifest.prompt,
+    expectedObservations: [...manifest.expectedObservations],
+  }));
+}
diff --git a/packages/paperclip-runner/src/mock-core/phase4b-demo-server.ts b/packages/paperclip-runner/src/mock-core/phase4b-demo-server.ts
index 88ff70132be..d3f03e81727 100644
--- a/packages/paperclip-runner/src/mock-core/phase4b-demo-server.ts
+++ b/packages/paperclip-runner/src/mock-core/phase4b-demo-server.ts
@@ -33,15 +33,21 @@ const MAX_BROWSER_EVENTS = 4096;
 
 export interface Phase4bDemoServerOptions {
   workingDirectory: string;
-  driverFactory: (envelope: Phase4TaskEnvelope) => HarnessDriver;
+  driverFactory: (
+    envelope: Phase4TaskEnvelope,
+    manifestId?: string,
+  ) => HarnessDriver;
   host?: string;
   port?: number;
+  /** Demo-chat manifests the browser lists; the server owns the catalogue. */
+  manifests?: readonly unknown[];
 }
 
 interface DemoEntry {
   id: string;
   runId: string;
   normalizedSessionId: string;
+  manifestId: string | null;
   envelope: Phase4TaskEnvelope;
   driver: HarnessDriver;
   session: HarnessSession;
@@ -54,6 +60,8 @@ interface DemoEntry {
 interface BrowserCreateBody {
   objective?: string;
   message?: string;
+  manifest?: string;
+  startTurn?: boolean;
 }
 
 function record(value: unknown): Record<string, unknown> {
@@ -160,18 +168,7 @@ export class Phase4bDemoServer {
 
   async start(): Promise<{ host: string; port: number; url: string }> {
     if (this.#server !== null) throw new Error("Phase 4b demo server is already started");
-    this.#server = createServer((request, response) => {
-      void this.#handle(request, response).catch((error) => {
-        if (!response.headersSent) {
-          json(response, errorStatus(error), {
-            error: record(error).code ?? "invalid_request",
-            message: redactCodexDiagnostic(String(error)),
-          });
-        } else {
-          response.end();
-        }
-      });
-    });
+    this.#server = createServer(this.middleware());
     await new Promise<void>((resolve, reject) => {
       this.#server!.once("error", reject);
       this.#server!.listen(this.#options.port, this.#options.host, () => resolve());
@@ -187,6 +184,25 @@ export class Phase4bDemoServer {
     };
   }
 
+  /**
+   * The same handler `start()` binds, so an embedding dev server (the Vite
+   * plugin) can mount the identical routes without a second implementation.
+   */
+  middleware(): (request: IncomingMessage, response: ServerResponse) => void {
+    return (request, response) => {
+      void this.#handle(request, response).catch((error) => {
+        if (!response.headersSent) {
+          json(response, errorStatus(error), {
+            error: record(error).code ?? "invalid_request",
+            message: redactCodexDiagnostic(String(error)),
+          });
+        } else {
+          response.end();
+        }
+      });
+    };
+  }
+
   async close(): Promise<void> {
     const entries = [...this.#entries.values()];
     this.#entries.clear();
@@ -215,13 +231,20 @@ export class Phase4bDemoServer {
       });
       return;
     }
+    if (request.method === "GET" && parts.join("/") === "api/phase4b/manifests") {
+      json(response, 200, { manifests: this.#options.manifests ?? [] });
+      return;
+    }
     if (request.method === "POST" && parts.join("/") === "api/phase4b/sessions") {
       const body = await readJson(request) as BrowserCreateBody;
       const objective = typeof body.objective === "string" && body.objective.trim().length > 0
         ? body.objective.trim()
         : "Complete the Phase 4b demo task safely.";
       const message = typeof body.message === "string" ? body.message : objective;
-      const entry = await this.#create(objective, message);
+      const manifestId = typeof body.manifest === "string" && body.manifest.length > 0
+        ? body.manifest
+        : null;
+      const entry = await this.#create(objective, message, manifestId, body.startTurn !== false);
       json(response, 201, await this.#publicState(entry));
       return;
     }
@@ -257,6 +280,10 @@ export class Phase4bDemoServer {
         "cache-control": "no-store",
         connection: "keep-alive",
       });
+      // Flush the headers immediately. A resumed subscriber can start past the
+      // last event, and Node would otherwise hold the response open with no
+      // bytes sent, leaving the browser stuck in CONNECTING.
+      response.write(": stream open\n\n");
       for (const event of entry.events.slice(after)) {
         response.write(`data: ${JSON.stringify(browserSafe(event))}\n\n`);
       }
@@ -269,7 +296,12 @@ export class Phase4bDemoServer {
       return;
     }
     const body = await readJson(request);
-    if (action === "steer") {
+    if (action === "turns") {
+      await entry.session.startTurn({
+        message: { role: "user", text: String(body.text ?? "") },
+      });
+      await new Promise<void>((resolve) => setImmediate(resolve));
+    } else if (action === "steer") {
       await entry.session.steer?.({
         turnId: String(body.turnId ?? ""),
         message: { role: "user", text: String(body.text ?? "") },
@@ -296,13 +328,18 @@ export class Phase4bDemoServer {
     json(response, 200, await this.#publicState(entry));
   }
 
-  async #create(objective: string, message: string): Promise<DemoEntry> {
+  async #create(
+    objective: string,
+    message: string,
+    manifestId: string | null = null,
+    startTurn = true,
+  ): Promise<DemoEntry> {
     const envelope = createPhase4TaskEnvelope({
       objective,
       contractRevision: "phase4b-demo-v1",
       criteria: [{ id: "objective", requirement: objective }],
     });
-    const driver = this.#options.driverFactory(envelope);
+    const driver = this.#options.driverFactory(envelope, manifestId ?? undefined);
     const id = randomUUID();
     const runId = `phase4b-run-${id}`;
     const normalizedSessionId = `phase4b-session-${id}`;
@@ -316,6 +353,7 @@ export class Phase4bDemoServer {
       id,
       runId,
       normalizedSessionId,
+      manifestId,
       envelope,
       driver,
       session,
@@ -326,8 +364,10 @@ export class Phase4bDemoServer {
     };
     this.#entries.set(id, entry);
     entry.consumeTask = this.#consume(entry, session);
-    await session.startTurn({ message: { role: "user", text: message } });
-    await new Promise<void>((resolve) => setImmediate(resolve));
+    if (startTurn) {
+      await session.startTurn({ message: { role: "user", text: message } });
+      await new Promise<void>((resolve) => setImmediate(resolve));
+    }
     return entry;
   }
 
@@ -383,6 +423,7 @@ export class Phase4bDemoServer {
       sessionId: entry.id,
       runId: entry.runId,
       normalizedSessionId: entry.normalizedSessionId,
+      manifest: entry.manifestId,
       providerAuthentication: "server-side",
       credentialsExposed: false,
       capabilities: entry.capabilities,
diff --git a/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.test.ts b/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.test.ts
new file mode 100644
index 00000000000..4dc04bdcea3
--- /dev/null
+++ b/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.test.ts
@@ -0,0 +1,331 @@
+import { describe, expect, it } from "vitest";
+
+import type { HarnessSession } from "../contracts/harness-driver.js";
+import { validatePrpEvent, type PrpEvent } from "../protocol/phase1-contract.js";
+import {
+  applyPrpEvent,
+  createSessionSnapshotFromMetadata,
+  type SessionSnapshot,
+} from "../reducer/session-reducer.js";
+import {
+  PHASE4B_DEMO_MANIFESTS,
+  findPhase4bDemoManifest,
+  phase4bDemoManifestCatalogue,
+} from "./phase4b-demo-manifests.js";
+import { Phase4bScriptedDriver } from "./phase4b-scripted-driver.js";
+
+/** Lets the collector loop drain events the driver already pushed. */
+async function tick(): Promise<void> {
+  await new Promise<void>((resolve) => setImmediate(resolve));
+}
+
+function collector(session: HarnessSession): {
+  events: PrpEvent[];
+  waitFor: (eventType: string, timeoutMs?: number) => Promise<PrpEvent>;
+  done: Promise<void>;
+} {
+  const events: PrpEvent[] = [];
+  const waiters: Array<{ eventType: string; resolve: (event: PrpEvent) => void }> = [];
+  const done = (async () => {
+    for await (const event of session.events()) {
+      events.push(event);
+      for (let index = waiters.length - 1; index >= 0; index -= 1) {
+        if (waiters[index]!.eventType === event.eventType) {
+          waiters.splice(index, 1)[0]!.resolve(event);
+        }
+      }
+    }
+  })();
+  return {
+    events,
+    done,
+    waitFor(eventType, timeoutMs = 5_000) {
+      const existing = events.find((event) => event.eventType === eventType);
+      if (existing !== undefined) return Promise.resolve(existing);
+      return new Promise<PrpEvent>((resolve, reject) => {
+        const timer = setTimeout(
+          () => reject(new Error(`timed out waiting for ${eventType}`)),
+          timeoutMs,
+        );
+        waiters.push({
+          eventType,
+          resolve: (event) => {
+            clearTimeout(timer);
+            resolve(event);
+          },
+        });
+      });
+    },
+  };
+}
+
+async function openSession(manifestId: string, chunkDelayMs = 0) {
+  const driver = new Phase4bScriptedDriver({ manifestId, chunkDelayMs });
+  const session = await driver.openSession({
+    runId: `run-${manifestId}`,
+    normalizedSessionId: `session-${manifestId}`,
+    workingDirectory: "/demo",
+  });
+  return { driver, session, stream: collector(session) };
+}
+
+function reduce(events: readonly PrpEvent[]): SessionSnapshot {
+  return events.reduce(
+    applyPrpEvent,
+    createSessionSnapshotFromMetadata({
+      fixtureName: "phase4b-demo",
+      identity: {
+        schema: "paperclip.prp.identity.v1",
+        companyId: "package-local-demo",
+        issueId: "phase4b-demo",
+        runId: events[0]?.runId ?? "run",
+        environmentLeaseId: "package-local",
+        runnerInstanceId: "phase4b-demo-server",
+        normalizedSessionId: events[0]?.normalizedSessionId ?? "session",
+      },
+      capabilities: {
+        schema: "paperclip.prp.capabilities.v1",
+        sessionReusePolicy: "reuse_per_issue",
+        driver: { kind: "phase4b-demo", version: "1" },
+        steer: true,
+        interrupt: true,
+        resume: true,
+        runtimeRequests: true,
+        structuredResult: true,
+        typedEvents: true,
+      },
+    }),
+  );
+}
+
+describe("Phase 4b demo manifests", () => {
+  it("gives every manifest a unique id and at least one expected observation", () => {
+    const ids = PHASE4B_DEMO_MANIFESTS.map((manifest) => manifest.id);
+    expect(new Set(ids).size).toBe(ids.length);
+    for (const manifest of PHASE4B_DEMO_MANIFESTS) {
+      expect(manifest.expectedObservations.length).toBeGreaterThan(0);
+      expect(manifest.prompt.length).toBeGreaterThan(0);
+    }
+  });
+
+  it("publishes a catalogue that carries no beat scripts to the browser", () => {
+    const catalogue = phase4bDemoManifestCatalogue();
+    expect(catalogue).toHaveLength(PHASE4B_DEMO_MANIFESTS.length);
+    expect(JSON.stringify(catalogue)).not.toContain("beats");
+  });
+
+  it("rejects an unknown manifest id", () => {
+    expect(() => findPhase4bDemoManifest("nope")).toThrow(/unknown Phase 4b demo manifest/);
+  });
+});
+
+describe("Phase 4b scripted driver", () => {
+  it("emits schema-valid canonical events for a completed turn", async () => {
+    const { session, stream } = await openSession("completion");
+    await session.startTurn({ message: { role: "user", text: "Summarize." } });
+    await stream.waitFor("turn.completed");
+
+    for (const event of stream.events) {
+      expect(validatePrpEvent(event).ok).toBe(true);
+    }
+    const snapshot = reduce(stream.events);
+    expect(snapshot.integrity).toBe("complete");
+    expect(snapshot.turnState).toBe("completed");
+    expect(snapshot.proposedResult?.reportedWorkDisposition).toBe("done");
+    expect(snapshot.items.map((item) => item.kind)).toEqual([
+      "reasoning",
+      "commandExecution",
+      "agentMessage",
+    ]);
+    const submitted = stream.events.find((event) => event.eventType === "turn.submitted");
+    expect((submitted?.payload as { text?: string }).text).toBe("Summarize.");
+    await session.close({ reason: "test" });
+  });
+
+  it("acknowledges same-turn steering and rejects a stale turn", async () => {
+    const { session, stream } = await openSession("steering", 20);
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Go." } });
+    await stream.waitFor("turn.started");
+    await session.steer?.({ turnId, message: { role: "user", text: "Be brief." } });
+
+    await tick();
+    const ack = stream.events.find(
+      (event) => (event.payload as { kind?: string }).kind === "steering_acknowledgement",
+    );
+    expect((ack?.payload as { text?: string }).text).toBe("Be brief.");
+
+    await expect(
+      session.steer?.({ turnId: "turn-999", message: { role: "user", text: "Stale." } }),
+    ).rejects.toMatchObject({ code: "stale_turn" });
+    expect(
+      stream.events.some(
+        (event) => (event.payload as { code?: string }).code === "stale_turn_rejected",
+      ),
+    ).toBe(true);
+
+    await stream.waitFor("turn.completed");
+    await expect(
+      session.steer?.({ turnId, message: { role: "user", text: "Too late." } }),
+    ).rejects.toMatchObject({ code: "already_terminal" });
+    await session.close({ reason: "test" });
+  });
+
+  it("cancels a turn interrupted before the provider accepts it", async () => {
+    const { session, stream } = await openSession("interrupt-before-start");
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Slow." } });
+    await session.interrupt?.({ turnId, reason: "browser_operator" });
+    await stream.waitFor("turn.cancelled");
+
+    expect(stream.events.some((event) => event.eventType === "turn.started")).toBe(false);
+    expect(reduce(stream.events).turnState).toBe("cancelled");
+    await session.close({ reason: "test" });
+  });
+
+  it("keeps partial text when a streaming item is interrupted", async () => {
+    const { session, stream } = await openSession("interrupt-during-generation", 40);
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Long." } });
+    await stream.waitFor("item.delta");
+    await session.interrupt?.({ turnId });
+    await stream.waitFor("turn.interrupted");
+
+    const snapshot = reduce(stream.events);
+    expect(snapshot.turnState).toBe("interrupted");
+    const item = snapshot.items.find((candidate) => candidate.itemId === "assistant-1");
+    expect(item?.text.length).toBeGreaterThan(0);
+    expect(item?.text).not.toContain("And this final one.");
+    await session.close({ reason: "test" });
+  });
+
+  it("keeps an interrupted tool item without fabricating a result", async () => {
+    const { session, stream } = await openSession("interrupt-during-tool", 5);
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Tool." } });
+    await stream.waitFor("item.started");
+    await session.interrupt?.({ turnId });
+    const interrupted = await stream.waitFor("turn.interrupted");
+    expect(interrupted.eventType).toBe("turn.interrupted");
+
+    const completion = stream.events.find(
+      (event) => event.eventType === "item.completed" && event.itemId === "tool-1",
+    );
+    expect((completion?.payload as { status?: string }).status).toBe("interrupted");
+    expect(completion?.payload).not.toHaveProperty("output");
+    await session.close({ reason: "test" });
+  });
+
+  it("resolves each runtime request exactly once", async () => {
+    const { session, stream } = await openSession("approvals");
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Approve." } });
+    const created = await stream.waitFor("runtime_request.created");
+    const request = (created.payload as { request: { requestId: string; actions: string[] } }).request;
+    expect(request.actions).toContain("accept_for_session");
+
+    await session.resolveRuntimeRequest?.({
+      requestId: request.requestId,
+      turnId,
+      resolution: { action: "accept" },
+    });
+    await expect(
+      session.resolveRuntimeRequest?.({
+        requestId: request.requestId,
+        turnId,
+        resolution: { action: "decline" },
+      }),
+    ).rejects.toMatchObject({ code: "already_terminal" });
+
+    await tick();
+    const resolved = stream.events.find(
+      (event) => event.eventType === "runtime_request.resolved",
+    );
+    expect((resolved?.payload as { action?: string }).action).toBe("accept");
+    await session.close({ reason: "test" });
+  });
+
+  it("never puts a submitted answer into the resolution acknowledgement", async () => {
+    const { session, stream } = await openSession("user-input");
+    const { turnId } = await session.startTurn({ message: { role: "user", text: "Ask." } });
+    const created = await stream.waitFor("runtime_request.created");
+    const request = (created.payload as { request: { requestId: string } }).request;
+    await session.resolveRuntimeRequest?.({
+      requestId: request.requestId,
+      turnId,
+      resolution: { action: "submit", answers: { environment: { answers: ["staging"] } } },
+    });
+    await tick();
+    const resolved = stream.events.find((event) => event.eventType === "runtime_request.resolved");
+    expect(JSON.stringify(resolved?.payload)).not.toContain("staging");
+    await session.close({ reason: "test" });
+  });
+
+  it("reports the exact diagnostic when goals are unsupported", async () => {
+    const { session } = await openSession("goals-unsupported");
+    await expect(session.goal?.({ action: "get" })).rejects.toThrow(
+      /does not advertise the goals capability/,
+    );
+    await session.close({ reason: "test" });
+  });
+
+  it("moves a goal through set, pause, resume, and clear", async () => {
+    const { session, stream } = await openSession("goals");
+    expect(await session.goal?.({ action: "get" })).toBeNull();
+    expect((await session.goal?.({ action: "set", objective: "Ship it" }))?.status).toBe("active");
+    expect((await session.goal?.({ action: "pause" }))?.status).toBe("paused");
+    expect((await session.goal?.({ action: "resume" }))?.status).toBe("active");
+    expect(await session.goal?.({ action: "clear" })).toBeNull();
+
+    await tick();
+    const goalEvents = stream.events.filter((event) =>
+      ["goal", "goal_cleared"].includes((event.payload as { kind?: string }).kind ?? ""),
+    );
+    expect(goalEvents).toHaveLength(4);
+    await session.close({ reason: "test" });
+  });
+
+  it("tracks child threads and keeps their terminal states", async () => {
+    const { session, stream } = await openSession("subagents");
+    await session.startTurn({ message: { role: "user", text: "Delegate." } });
+    await stream.waitFor("turn.completed");
+
+    const lineage = session.lineage?.() ?? [];
+    expect(lineage.map((entry) => entry.threadId)).toEqual([
+      "thread-session-subagents",
+      "thread-child-a",
+      "thread-child-b",
+    ]);
+    expect(lineage[1]?.status).toBe("completed");
+    expect(lineage[2]?.status).toBe("failed");
+    await session.close({ reason: "test" });
+  });
+
+  it("reports a failed turn and a blocked disposition", async () => {
+    const { session, stream } = await openSession("failure");
+    await session.startTurn({ message: { role: "user", text: "Fail." } });
+    await stream.waitFor("turn.failed");
+
+    const snapshot = reduce(stream.events);
+    expect(snapshot.turnState).toBe("failed");
+    expect(snapshot.items[0]?.status).toBe("failed");
+    expect(snapshot.proposedResult?.reportedWorkDisposition).toBe("blocked");
+    await session.close({ reason: "test" });
+  });
+
+  it("resumes the same session identity and continues the source sequence", async () => {
+    const { driver, session, stream } = await openSession("completion");
+    await session.startTurn({ message: { role: "user", text: "First." } });
+    await stream.waitFor("turn.completed");
+    const persisted = await session.snapshot();
+    await session.close({ reason: "browser_reconnect" });
+
+    const recovery = await driver.recoverSession(persisted);
+    expect(recovery.recovered).toBe(true);
+    const resumed = recovery.session!;
+    const resumedStream = collector(resumed);
+    await resumedStream.waitFor("session.resumed");
+
+    const combined = [...stream.events, ...resumedStream.events];
+    const snapshot = reduce(combined);
+    expect(snapshot.integrity).toBe("complete");
+    expect(snapshot.identity.normalizedSessionId).toBe("session-completion");
+    expect(resumed.ids().driverSessionId).toBe(session.ids().driverSessionId);
+    await resumed.close({ reason: "test" });
+  });
+});
diff --git a/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.ts b/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.ts
new file mode 100644
index 00000000000..1a511c9a0a4
--- /dev/null
+++ b/packages/paperclip-runner/src/mock-core/phase4b-scripted-driver.ts
@@ -0,0 +1,756 @@
+import {
+  HarnessCapabilityUnavailableError,
+  HarnessOperationAlreadyTerminalError,
+  HarnessStaleTurnError,
+  type HarnessDriver,
+  type HarnessDriverDescriptor,
+  type HarnessGoalOperation,
+  type HarnessRuntimeRequest,
+  type HarnessRuntimeRequestResolution,
+  type HarnessSession,
+  type HarnessSessionRecoveryResult,
+  type HarnessThreadGoal,
+  type HarnessThreadLineageEntry,
+  type OpenHarnessSessionInput,
+  type PersistedHarnessSession,
+} from "../contracts/harness-driver.js";
+import type { NativeUserMessage } from "../contracts/types.js";
+import type {
+  PrpEvent,
+  PrpStructuredRunResult,
+} from "../protocol/phase1-contract.js";
+import {
+  findPhase4bDemoManifest,
+  type Phase4bDemoBeat,
+  type Phase4bDemoManifest,
+} from "./phase4b-demo-manifests.js";
+
+const DRIVER_KIND = "phase4b-scripted-demo";
+
+/**
+ * `setTimeout` is typed as a DOM timer when this module is compiled under the
+ * browser tsconfig for the transcript-model tests, so `unref` is guarded
+ * rather than assumed.
+ */
+function unrefTimer(timer: ReturnType<typeof setTimeout>): void {
+  const handle = timer as unknown as { unref?: () => void };
+  if (typeof handle.unref === "function") handle.unref();
+}
+
+const HOLD_TIMEOUT_MS = 45_000;
+
+export interface Phase4bScriptedDriverOptions {
+  manifestId?: string;
+  manifest?: Phase4bDemoManifest;
+  /** Milliseconds between streamed chunks. Tests set this to zero. */
+  chunkDelayMs?: number;
+  now?: () => Date;
+}
+
+class EventQueue implements AsyncIterable<PrpEvent> {
+  #values: PrpEvent[] = [];
+  #waiters: Array<(result: IteratorResult<PrpEvent>) => void> = [];
+  #closed = false;
+
+  push(value: PrpEvent): void {
+    if (this.#closed) return;
+    const waiter = this.#waiters.shift();
+    if (waiter !== undefined) waiter({ value, done: false });
+    else this.#values.push(value);
+  }
+
+  close(): void {
+    this.#closed = true;
+    for (const waiter of this.#waiters.splice(0)) {
+      waiter({ value: undefined, done: true });
+    }
+  }
+
+  [Symbol.asyncIterator](): AsyncIterator<PrpEvent> {
+    return {
+      next: async () => {
+        const value = this.#values.shift();
+        if (value !== undefined) return { value, done: false };
+        if (this.#closed) return { value: undefined, done: true };
+        return new Promise<IteratorResult<PrpEvent>>((resolve) => {
+          this.#waiters.push(resolve);
+        });
+      },
+    };
+  }
+}
+
+/** A promise that a later event can settle early. */
+class Gate {
+  #resolve: (() => void) | null = null;
+  #settled = false;
+  readonly promise: Promise<void>;
+
+  constructor() {
+    this.promise = new Promise<void>((resolve) => {
+      this.#resolve = resolve;
+    });
+  }
+
+  open(): void {
+    if (this.#settled) return;
+    this.#settled = true;
+    this.#resolve?.();
+  }
+
+  get settled(): boolean {
+    return this.#settled;
+  }
+}
+
+interface PendingRequestEntry {
+  request: HarnessRuntimeRequest;
+  gate: Gate;
+  outcome: "pending" | "resolved" | "expired" | "cancelled";
+  action: string | null;
+  timer: ReturnType<typeof setTimeout> | null;
+}
+
+interface ScriptedSessionInput extends OpenHarnessSessionInput {
+  manifest: Phase4bDemoManifest;
+  chunkDelayMs: number;
+  now: () => Date;
+  resumed: boolean;
+  restoredSequence: number;
+  restoredGoal: HarnessThreadGoal | null;
+  restoredLineage: HarnessThreadLineageEntry[];
+  turnCounter: number;
+}
+
+class Phase4bScriptedSession implements HarnessSession {
+  readonly #queue = new EventQueue();
+  readonly #input: ScriptedSessionInput;
+  readonly #pending = new Map<string, PendingRequestEntry>();
+  readonly #lineage: HarnessThreadLineageEntry[];
+  readonly #terminalTurns: Array<{ turnId: string; fingerprint: string }> = [];
+  #sequence: number;
+  #turnCounter: number;
+  #activeTurnId: string | null = null;
+  #turnStartPending = false;
+  #interruptGate: Gate | null = null;
+  #goal: HarnessThreadGoal | null;
+  #semanticResult: PrpStructuredRunResult | null = null;
+  #closed = false;
+  #turnRunner: Promise<void> = Promise.resolve();
+
+  constructor(input: ScriptedSessionInput) {
+    this.#input = input;
+    this.#sequence = input.restoredSequence;
+    this.#turnCounter = input.turnCounter;
+    this.#goal = input.restoredGoal;
+    this.#lineage = [
+      {
+        threadId: this.#threadId,
+        providerSessionId: input.normalizedSessionId,
+        parentThreadId: null,
+        depth: 0,
+        nickname: "root",
+        role: "primary",
+        status: "running",
+      },
+      ...input.restoredLineage.filter((entry) => entry.depth > 0),
+    ];
+    this.#emit(input.resumed ? "session.resumed" : "session.started", {
+      driverSessionId: this.#threadId,
+      providerSessionId: input.normalizedSessionId,
+      manifest: input.manifest.id,
+      workingDirectory: input.workingDirectory,
+      providerAuthentication: "server-side",
+    });
+  }
+
+  get #threadId(): string {
+    return `thread-${this.#input.normalizedSessionId}`;
+  }
+
+  ids(): { driverSessionId: string; providerSessionId?: string | null; displayId?: string | null } {
+    return {
+      driverSessionId: this.#threadId,
+      providerSessionId: this.#input.normalizedSessionId,
+      displayId: this.#input.manifest.name,
+    };
+  }
+
+  events(): AsyncIterable<PrpEvent> {
+    return this.#queue;
+  }
+
+  async startTurn(input: { message: NativeUserMessage }): Promise<{ turnId: string }> {
+    if (this.#closed) throw new Error("session is closed");
+    if (this.#activeTurnId !== null || this.#turnStartPending) {
+      throw new Error("a turn is already active");
+    }
+    this.#turnCounter += 1;
+    const turnId = `turn-${this.#turnCounter}`;
+    const beats = this.#turnCounter === 1
+      ? this.#input.manifest.beats
+      : (this.#input.manifest.followUpBeats ?? this.#input.manifest.beats);
+    this.#turnStartPending = true;
+    this.#interruptGate = new Gate();
+    this.#emit("turn.submitted", {
+      turnId,
+      text: input.message.text,
+      envelopeSchema: "paperclip.task_envelope.v1",
+    }, { turnId });
+    this.#turnRunner = this.#runTurn(turnId, beats).catch((error) => {
+      this.#emit("harness.diagnostic", { message: `demo turn failed: ${String(error)}` });
+    });
+    return { turnId };
+  }
+
+  async steer(input: { turnId: string; message: NativeUserMessage }): Promise<void> {
+    if (!this.#input.manifest.capabilities.steering) {
+      throw new HarnessCapabilityUnavailableError(
+        "steering",
+        "this demo manifest does not advertise steering",
+      );
+    }
+    if (this.#activeTurnId === null && !this.#turnStartPending) {
+      throw new HarnessOperationAlreadyTerminalError("steer");
+    }
+    if (input.turnId !== this.#activeTurnId) {
+      this.#emit("harness.diagnostic", {
+        code: "stale_turn_rejected",
+        message: `steering rejected: turn ${input.turnId} is not the active turn`,
+      });
+      throw new HarnessStaleTurnError(input.turnId);
+    }
+    this.#emit("item.completed", {
+      kind: "steering_acknowledgement",
+      status: "acknowledged",
+      text: input.message.text,
+    }, { turnId: input.turnId, itemId: `steer-${input.turnId}-${this.#sequence + 1}` });
+  }
+
+  async interrupt(input: { turnId?: string; reason?: string }): Promise<void> {
+    if (!this.#input.manifest.capabilities.interruption) {
+      throw new HarnessCapabilityUnavailableError(
+        "interruption",
+        "this demo manifest does not advertise interruption",
+      );
+    }
+    if (this.#activeTurnId === null && !this.#turnStartPending) {
+      throw new HarnessOperationAlreadyTerminalError("interrupt");
+    }
+    if (
+      input.turnId !== undefined &&
+      this.#activeTurnId !== null &&
+      input.turnId !== this.#activeTurnId
+    ) {
+      throw new HarnessStaleTurnError(input.turnId);
+    }
+    this.#emit("item.completed", {
+      kind: "interrupt_acknowledgement",
+      status: "acknowledged",
+      text: input.reason ?? "browser_operator",
+    }, {
+      ...(this.#activeTurnId === null ? {} : { turnId: this.#activeTurnId }),
+      itemId: `interrupt-${this.#sequence + 1}`,
+    });
+    this.#interruptGate?.open();
+  }
+
+  pendingRuntimeRequests(): HarnessRuntimeRequest[] {
+    return [...this.#pending.values()]
+      .filter((entry) => entry.outcome === "pending")
+      .map((entry) => structuredClone(entry.request));
+  }
+
+  async resolveRuntimeRequest(input: {
+    requestId: string;
+    turnId: string;
+    resolution: HarnessRuntimeRequestResolution;
+  }): Promise<void> {
+    const entry = this.#pending.get(input.requestId);
+    if (entry === undefined) {
+      throw new Error(`unknown runtime request ${input.requestId}`);
+    }
+    if (entry.outcome !== "pending") {
+      throw new HarnessOperationAlreadyTerminalError(`resolve ${input.requestId}`);
+    }
+    if (input.turnId.length > 0 && input.turnId !== entry.request.turnId) {
+      throw new HarnessStaleTurnError(input.turnId);
+    }
+    entry.outcome = "resolved";
+    entry.action = input.resolution.action;
+    this.#settleRequest(entry, "runtime_request.resolved");
+  }
+
+  async goal(input: HarnessGoalOperation): Promise<HarnessThreadGoal | null> {
+    if (this.#input.manifest.capabilities.goals !== true) {
+      throw new HarnessCapabilityUnavailableError(
+        "goal",
+        this.#input.manifest.capabilities.unsupported?.[0] ??
+          "this app-server does not advertise the goals capability",
+      );
+    }
+    const at = this.#input.now().getTime();
+    if (input.action === "get") return this.#goal === null ? null : structuredClone(this.#goal);
+    if (input.action === "set") {
+      this.#goal = {
+        threadId: this.#threadId,
+        objective: input.objective,
+        status: "active",
+        tokenBudget: input.tokenBudget ?? null,
+        tokensUsed: 0,
+        timeUsedSeconds: 0,
+        createdAt: this.#goal?.createdAt ?? at,
+        updatedAt: at,
+      };
+    } else if (input.action === "clear") {
+      this.#goal = null;
+    } else if (this.#goal === null) {
+      throw new HarnessCapabilityUnavailableError(
+        `goal ${input.action}`,
+        "no goal is set on this thread",
+      );
+    } else {
+      this.#goal = {
+        ...this.#goal,
+        status: input.action === "pause" ? "paused" : "active",
+        updatedAt: at,
+      };
+    }
+    this.#emit("item.completed", {
+      kind: this.#goal === null ? "goal_cleared" : "goal",
+      text: this.#goal === null
+        ? "Goal cleared."
+        : `Goal ${input.action === "set" ? "set" : input.action}: ${this.#goal.objective}`,
+      operation: input.action,
+      ...(this.#goal === null ? {} : { goal: structuredClone(this.#goal) }),
+    }, { itemId: `goal-${this.#sequence + 1}` });
+    return this.#goal === null ? null : structuredClone(this.#goal);
+  }
+
+  lineage(): HarnessThreadLineageEntry[] {
+    return structuredClone(this.#lineage);
+  }
+
+  async usage(): Promise<Record<string, unknown> | null> {
+    return { inputTokens: this.#sequence * 12, outputTokens: this.#sequence * 7 };
+  }
+
+  async snapshot(): Promise<PersistedHarnessSession> {
+    return {
+      driverKind: DRIVER_KIND,
+      driverSessionId: this.#threadId,
+      providerSessionId: this.#input.normalizedSessionId,
+      runId: this.#input.runId,
+      normalizedSessionId: this.#input.normalizedSessionId,
+      activeTurnId: this.#activeTurnId,
+      ...(this.#semanticResult === null
+        ? {}
+        : {
+            semanticResult: {
+              result: structuredClone(this.#semanticResult),
+              fingerprint: `${this.#input.manifest.id}:${this.#turnCounter}`,
+              turnId: `turn-${this.#turnCounter}`,
+            },
+          }),
+      terminalTurns: structuredClone(this.#terminalTurns),
+      pendingRuntimeRequests: this.pendingRuntimeRequests(),
+      goal: this.#goal === null ? null : structuredClone(this.#goal),
+      lineage: this.lineage(),
+      lastSourceSequence: this.#sequence,
+    };
+  }
+
+  async close(input: { reason: string; force?: boolean }): Promise<void> {
+    if (this.#closed) return;
+    this.#closed = true;
+    for (const entry of this.#pending.values()) {
+      if (entry.outcome !== "pending") continue;
+      entry.outcome = "cancelled";
+      this.#settleRequest(entry, "runtime_request.cancelled");
+    }
+    this.#interruptGate?.open();
+    await this.#turnRunner.catch(() => undefined);
+    this.#emit("session.closed", { reason: input.reason });
+    this.#queue.close();
+  }
+
+  /** Test and reconnect seam: the demo server uses it to keep the cursor. */
+  get sequence(): number {
+    return this.#sequence;
+  }
+
+  get turnCounter(): number {
+    return this.#turnCounter;
+  }
+
+  #emit(
+    eventType: PrpEvent["eventType"],
+    payload: Record<string, unknown>,
+    binding: { turnId?: string; itemId?: string } = {},
+  ): void {
+    this.#sequence += 1;
+    this.#queue.push({
+      schema: "paperclip.prp.event.v1",
+      sourceEventId: `${this.#input.runId}:${this.#sequence}`,
+      sourceSeq: this.#sequence,
+      sourceInstanceId: `${DRIVER_KIND}:${this.#input.normalizedSessionId}`,
+      sourceKind: "runner",
+      runId: this.#input.runId,
+      normalizedSessionId: this.#input.normalizedSessionId,
+      ...(binding.turnId === undefined ? {} : { turnId: binding.turnId }),
+      ...(binding.itemId === undefined ? {} : { itemId: binding.itemId }),
+      eventType,
+      schemaVersion: 1,
+      priority: 1,
+      emittedAt: this.#input.now().toISOString(),
+      payload,
+    } as PrpEvent);
+  }
+
+  #settleRequest(entry: PendingRequestEntry, eventType: PrpEvent["eventType"]): void {
+    if (entry.timer !== null) {
+      clearTimeout(entry.timer);
+      entry.timer = null;
+    }
+    this.#emit(eventType, {
+      requestId: entry.request.requestId,
+      requestKind: entry.request.requestKind,
+      ...(entry.action === null ? {} : { action: entry.action }),
+    }, { turnId: entry.request.turnId, itemId: entry.request.itemId });
+    entry.gate.open();
+  }
+
+  async #sleep(ms: number): Promise<"elapsed" | "interrupted"> {
+    if (ms <= 0) return this.#interruptGate?.settled === true ? "interrupted" : "elapsed";
+    const gate = this.#interruptGate;
+    return await new Promise<"elapsed" | "interrupted">((resolve) => {
+      const timer = setTimeout(() => resolve("elapsed"), ms);
+      unrefTimer(timer);
+      void gate?.promise.then(() => {
+        clearTimeout(timer);
+        resolve("interrupted");
+      });
+    });
+  }
+
+  get #interrupted(): boolean {
+    return this.#interruptGate?.settled === true;
+  }
+
+  async #runTurn(turnId: string, beats: readonly Phase4bDemoBeat[]): Promise<void> {
+    const manifest = this.#input.manifest;
+    if (manifest.startDelayMs > 0) {
+      const outcome = await this.#sleep(manifest.startDelayMs);
+      if (outcome === "interrupted" || this.#closed) {
+        this.#turnStartPending = false;
+        this.#recordTerminalTurn(turnId, "cancelled");
+        this.#emit("turn.cancelled", {
+          turnId,
+          reason: "interrupted before the provider accepted the turn",
+        }, { turnId });
+        return;
+      }
+    }
+    if (this.#closed) {
+      this.#turnStartPending = false;
+      return;
+    }
+    this.#turnStartPending = false;
+    this.#activeTurnId = turnId;
+    this.#emit("turn.accepted", { turnId }, { turnId });
+    this.#emit("turn.started", { turnId, status: "inProgress" }, { turnId });
+    this.#emit("runtime.phase.changed", { phase: "harness_running" });
+
+    let failed = false;
+    for (const beat of beats) {
+      if (this.#interrupted || this.#closed) break;
+      failed = (await this.#runBeat(turnId, beat)) || failed;
+    }
+
+    this.#activeTurnId = null;
+    if (this.#interrupted) {
+      this.#recordTerminalTurn(turnId, "interrupted");
+      this.#emit("turn.interrupted", { turnId, reason: "browser_operator" }, { turnId });
+      return;
+    }
+    if (this.#closed) return;
+    const disposition = failed || manifest.disposition === "failed" ? "failed" : "completed";
+    this.#recordTerminalTurn(turnId, disposition);
+    this.#emitResult(turnId, disposition === "failed");
+    this.#emit(disposition === "failed" ? "turn.failed" : "turn.completed", {
+      turnId,
+      ...(disposition === "failed"
+        ? { message: "The demo manifest ends in a failed turn." }
+        : {}),
+    }, { turnId });
+  }
+
+  #recordTerminalTurn(turnId: string, state: string): void {
+    this.#terminalTurns.push({ turnId, fingerprint: `${state}:${turnId}` });
+  }
+
+  async #runBeat(turnId: string, beat: Phase4bDemoBeat): Promise<boolean> {
+    const delay = this.#input.chunkDelayMs;
+    if (beat.kind === "reasoning" || beat.kind === "assistant") {
+      const itemKind = beat.kind === "reasoning" ? "reasoning" : "agentMessage";
+      this.#emit("item.started", { kind: itemKind, text: "" }, { turnId, itemId: beat.itemId });
+      let text = "";
+      for (const chunk of beat.chunks) {
+        if ((await this.#sleep(delay)) === "interrupted" || this.#closed) {
+          this.#emit("item.completed", {
+            kind: itemKind,
+            text,
+            status: "interrupted",
+            interrupted: true,
+          }, { turnId, itemId: beat.itemId });
+          return false;
+        }
+        text += chunk;
+        this.#emit("item.delta", { kind: itemKind, text: chunk }, { turnId, itemId: beat.itemId });
+      }
+      this.#emit("item.completed", { kind: itemKind, text, status: "completed" }, {
+        turnId,
+        itemId: beat.itemId,
+      });
+      return false;
+    }
+
+    if (beat.kind === "tool") {
+      this.#emit("item.started", {
+        kind: "commandExecution",
+        text: `${beat.name}: ${beat.input}`,
+        toolName: beat.name,
+        input: beat.input,
+        status: "running",
+      }, { turnId, itemId: beat.itemId });
+      const wait = beat.hold === true ? HOLD_TIMEOUT_MS : Math.max(delay, 0) * 2;
+      if ((await this.#sleep(wait)) === "interrupted" || this.#closed) {
+        this.#emit("item.completed", {
+          kind: "commandExecution",
+          text: `${beat.name}: ${beat.input}`,
+          toolName: beat.name,
+          input: beat.input,
+          status: "interrupted",
+          interrupted: true,
+        }, { turnId, itemId: beat.itemId });
+        return false;
+      }
+      if (beat.failure !== undefined) {
+        this.#emit("item.failed", {
+          kind: "commandExecution",
+          text: `${beat.name}: ${beat.input}`,
+          toolName: beat.name,
+          input: beat.input,
+          message: beat.failure,
+          status: "failed",
+        }, { turnId, itemId: beat.itemId });
+        return true;
+      }
+      this.#emit("item.completed", {
+        kind: "commandExecution",
+        text: `${beat.name}: ${beat.input}`,
+        toolName: beat.name,
+        input: beat.input,
+        output: beat.output ?? "",
+        status: "completed",
+      }, { turnId, itemId: beat.itemId });
+      return false;
+    }
+
+    if (beat.kind === "request") {
+      const request: HarnessRuntimeRequest = {
+        requestId: beat.requestId,
+        requestKind: beat.requestKind,
+        method: beat.method,
+        turnId,
+        itemId: beat.itemId,
+        status: "pending",
+        prompt: beat.prompt,
+        details: structuredClone(beat.details),
+      };
+      const entry: PendingRequestEntry = {
+        request,
+        gate: new Gate(),
+        outcome: "pending",
+        action: null,
+        timer: null,
+      };
+      this.#pending.set(request.requestId, entry);
+      if (beat.expiresAfterMs !== undefined) {
+        entry.timer = setTimeout(() => {
+          if (entry.outcome !== "pending") return;
+          entry.outcome = "expired";
+          this.#settleRequest(entry, "runtime_request.expired");
+        }, beat.expiresAfterMs);
+        unrefTimer(entry.timer);
+      }
+      this.#emit("runtime_request.created", {
+        request: {
+          ...request,
+          type: request.method,
+          actions: [...beat.actions],
+        },
+      }, { turnId, itemId: beat.itemId });
+      await Promise.race([
+        entry.gate.promise,
+        this.#interruptGate?.promise ?? new Promise<void>(() => undefined),
+      ]);
+      if (entry.outcome === "pending") {
+        entry.outcome = "cancelled";
+        this.#settleRequest(entry, "runtime_request.cancelled");
+        return false;
+      }
+      return false;
+    }
+
+    if (beat.kind === "child") {
+      const child: HarnessThreadLineageEntry = {
+        threadId: beat.threadId,
+        providerSessionId: `${beat.threadId}-provider`,
+        parentThreadId: this.#threadId,
+        depth: 1,
+        nickname: beat.nickname,
+        role: beat.role,
+        status: "starting",
+      };
+      this.#lineage.push(child);
+      for (const status of beat.statuses) {
+        if ((await this.#sleep(delay)) === "interrupted" || this.#closed) return false;
+        child.status = status;
+        this.#emit("item.completed", {
+          kind: "thread_lineage",
+          text: `Child thread ${beat.nickname} is ${status}.`,
+          thread: structuredClone(child),
+        }, { turnId, itemId: `${beat.threadId}-${status}` });
+      }
+      return false;
+    }
+
+    if (beat.kind === "diagnostic") {
+      this.#emit("harness.diagnostic", { message: beat.message });
+      return true;
+    }
+
+    this.#emit("harness.diagnostic", { message: beat.note });
+    await this.#sleep(HOLD_TIMEOUT_MS);
+    return false;
+  }
+
+  #emitResult(turnId: string, failed: boolean): void {
+    const manifest = this.#input.manifest;
+    const result: PrpStructuredRunResult = {
+      schema: "paperclip.run_result.v1",
+      reportedWorkDisposition: failed ? "blocked" : "done",
+      summary: failed
+        ? `The ${manifest.name} demo ended in a failed turn on purpose.`
+        : `The ${manifest.name} demo turn completed.`,
+      completionClaim: {
+        contractRevision: "phase4b-demo-v1",
+        objectiveSatisfied: !failed,
+        criteria: [
+          {
+            criterionId: "objective",
+            status: failed ? "not_satisfied" : "satisfied",
+            evidenceRefs: [`manifest:${manifest.id}`],
+          },
+        ],
+        remainingWork: failed
+          ? [
+              {
+                description: "Investigate the scripted command failure.",
+                blocksCompletion: true,
+              },
+            ]
+          : [],
+      },
+      evidence: [{ ref: `manifest:${manifest.id}` }],
+      verification: [
+        {
+          commandOrCheck: `phase4b demo manifest ${manifest.id}`,
+          status: failed ? "failed" : "passed",
+        },
+      ],
+      attentionRequests: [],
+      artifacts: [],
+    };
+    this.#semanticResult = result;
+    this.#emit("run.result.proposed", result as unknown as Record<string, unknown>, { turnId });
+  }
+}
+
+/**
+ * A package-local `HarnessDriver` that replays a deterministic demo manifest.
+ * It exists so the browser tracer can reach every protocol state without a
+ * provider account, using exactly the routes the real Codex boundary uses.
+ */
+export class Phase4bScriptedDriver implements HarnessDriver {
+  readonly #manifest: Phase4bDemoManifest;
+  readonly #chunkDelayMs: number;
+  readonly #now: () => Date;
+  #session: Phase4bScriptedSession | null = null;
+
+  constructor(options: Phase4bScriptedDriverOptions = {}) {
+    this.#manifest = options.manifest ??
+      findPhase4bDemoManifest(options.manifestId ?? "completion");
+    this.#chunkDelayMs = options.chunkDelayMs ?? 45;
+    this.#now = options.now ?? (() => new Date());
+  }
+
+  get manifest(): Phase4bDemoManifest {
+    return this.#manifest;
+  }
+
+  async descriptor(): Promise<HarnessDriverDescriptor> {
+    return {
+      kind: DRIVER_KIND,
+      displayName: `Phase 4b demo — ${this.#manifest.name}`,
+      version: "1",
+      protocolVersion: "phase4b-demo-v1",
+      capabilities: structuredClone(this.#manifest.capabilities),
+    };
+  }
+
+  async openSession(input: OpenHarnessSessionInput): Promise<HarnessSession> {
+    const session = new Phase4bScriptedSession({
+      ...input,
+      manifest: this.#manifest,
+      chunkDelayMs: this.#chunkDelayMs,
+      now: this.#now,
+      resumed: false,
+      restoredSequence: 0,
+      restoredGoal: null,
+      restoredLineage: [],
+      turnCounter: 0,
+    });
+    this.#session = session;
+    return session;
+  }
+
+  async recoverSession(
+    snapshot: PersistedHarnessSession,
+  ): Promise<HarnessSessionRecoveryResult> {
+    if (snapshot.driverKind !== DRIVER_KIND) {
+      return { recovered: false, reason: `cannot resume driver kind ${snapshot.driverKind}` };
+    }
+    const runId = snapshot.runId;
+    const normalizedSessionId = snapshot.normalizedSessionId;
+    if (runId === undefined || normalizedSessionId === undefined) {
+      return { recovered: false, reason: "snapshot is missing its runner identity" };
+    }
+    const session = new Phase4bScriptedSession({
+      runId,
+      normalizedSessionId,
+      workingDirectory: "<server-owned>",
+      manifest: this.#manifest,
+      chunkDelayMs: this.#chunkDelayMs,
+      now: this.#now,
+      resumed: true,
+      restoredSequence: snapshot.lastSourceSequence ?? 0,
+      restoredGoal: snapshot.goal ?? null,
+      restoredLineage: snapshot.lineage ?? [],
+      turnCounter: this.#session?.turnCounter ?? 0,
+    });
+    this.#session = session;
+    return { recovered: true, session };
+  }
+}
diff --git a/packages/paperclip-runner/test-fixtures/forbidden-ui-runtime/devtools/browser/adapted-component.tsx b/packages/paperclip-runner/test-fixtures/forbidden-ui-runtime/devtools/browser/adapted-component.tsx
new file mode 100644
index 00000000000..3a9a1ad4641
--- /dev/null
+++ b/packages/paperclip-runner/test-fixtures/forbidden-ui-runtime/devtools/browser/adapted-component.tsx
@@ -0,0 +1,4 @@
+// Negative fixture: adapting AI Elements source must not pull the AI SDK.
+import { useChat } from "@ai-sdk/react";
+
+export const chat = useChat;
diff --git a/packages/paperclip-runner/vite.config.ts b/packages/paperclip-runner/vite.config.ts
index 6cbfced8ed0..1537c73ce70 100644
--- a/packages/paperclip-runner/vite.config.ts
+++ b/packages/paperclip-runner/vite.config.ts
@@ -5,12 +5,13 @@ import react from "@vitejs/plugin-react";
 import { defineConfig } from "vite";
 
 import { phase2BrowserServerPlugin } from "./scripts/phase2-browser-server.mjs";
+import { phase4bBrowserServerPlugin } from "./scripts/phase4b-browser-server.mjs";
 
 const packageRoot = fileURLToPath(new URL(".", import.meta.url));
 
 export default defineConfig({
   root: resolve(packageRoot, "devtools/browser"),
-  plugins: [react(), phase2BrowserServerPlugin()],
+  plugins: [react(), phase2BrowserServerPlugin(), phase4bBrowserServerPlugin()],
   server: {
     host: "127.0.0.1",
   },
diff --git a/packages/paperclip-runner/vitest.config.ts b/packages/paperclip-runner/vitest.config.ts
index 2180e29a45a..c0828d81567 100644
--- a/packages/paperclip-runner/vitest.config.ts
+++ b/packages/paperclip-runner/vitest.config.ts
@@ -3,6 +3,10 @@ import { defineConfig } from "vitest/config";
 export default defineConfig({
   test: {
     environment: "node",
-    include: ["src/**/*.test.ts", "scripts/phase2-browser-server.test.mjs"],
+    include: [
+      "src/**/*.test.ts",
+      "devtools/browser/src/**/*.test.ts",
+      "scripts/phase2-browser-server.test.mjs",
+    ],
   },
 });