diff --git a/change-logs/2026/06/27/fix-verify-task-create-persistence.md b/change-logs/2026/06/27/fix-verify-task-create-persistence.md
new file mode 100644
index 00000000..171e6050
--- /dev/null
+++ b/change-logs/2026/06/27/fix-verify-task-create-persistence.md
@@ -0,0 +1 @@
+Task creation now verifies the write actually landed on disk before reporting success. Previously `dev3 task create` could print "Created task <id>" (consuming a seq) for a task that was never persisted; it now fails loudly with a clear error instead of returning a ghost task, so the printed id is always immediately resolvable.
diff --git a/decisions/082-verify-task-create-persistence.md b/decisions/082-verify-task-create-persistence.md
new file mode 100644
index 00000000..368fa2c9
--- /dev/null
+++ b/decisions/082-verify-task-create-persistence.md
@@ -0,0 +1,37 @@
+# 082 — Verify task-create persistence before reporting success
+
+## Context
+Agent vents reported that `dev3 task create` printed `Created task <id> (seq N)` and exited 0,
+but the task never appeared in `dev3 tasks list` / `dev3 task show` — the seq was consumed yet
+the task was lost. An agent trusts that success line and tells the user "task created" when it
+wasn't.
+
+## Investigation
+The single-instance data layer is sound: `addTask` reads fresh under a cross-process
+`mkdir` file lock (`file-lock.ts`), writes via `atomicWriteFile` (temp + `rename`), and
+invalidates the read cache. No public `saveTasks(fullArray)` caller exists outside `data.ts`,
+so there is no stale read-modify-write clobber within one process. The read cache is
+stat-validated and adds always grow file size, so a stale cache cannot hide a freshly added
+task. The reproducible permanent-loss paths are out of scope here: multiple app instances
+clobbering the shared file (`wontfix` — niche dev-only setup) and macOS Full Disk Access loss
+mid-write (separate FDA issue). What we *can* fix is the dishonest success.
+
+## Decision
+`addTask` (`src/bun/data.ts`) now re-reads the tasks file fresh (`rawLoadTasks(..., { strict:
+true })`, bypassing the cache) after the save and throws if the new task id is absent. The error
+names the likely causes (FDA / multi-instance clobber) and the file path. Because the RPC layer
+turns a thrown handler error into `{ ok: false, error }`, the CLI now exits non-zero with a clear
+message instead of printing a false "Created task". This also guarantees the "id is immediately
+resolvable" contract: if create returns, the task is confirmed on disk.
+
+## Risks
+One extra fresh read per task creation (negligible — task creation is a rare, user/agent-driven
+action and the file is small). The guard does not catch a *later* clobber by another instance
+(that happens after `addTask` returns); it only guarantees the write landed at creation time.
+
+## Alternatives considered
+- Verify in the RPC handler `createTask` instead of the data layer — rejected: would only cover
+  the create RPC path, not every `addTask` caller (UI, variants).
+- Chase the multi-instance / FDA root cause — out of scope (wontfix / separate FDA track).
+- Block/retry until queryable — unnecessary; the write is synchronous under the lock, so a
+  single read-back is authoritative.
diff --git a/src/bun/__tests__/data-create-verify.test.ts b/src/bun/__tests__/data-create-verify.test.ts
new file mode 100644
index 00000000..09d36ee6
--- /dev/null
+++ b/src/bun/__tests__/data-create-verify.test.ts
@@ -0,0 +1,97 @@
+import { afterAll, afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+import { mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import type { Project, Task } from "../../shared/types";
+
+// Mock node:fs/promises so we can simulate a *silent write drop*: the payload
+// write "succeeds" (no throw) but the new task never actually lands on disk.
+// This is the failure mode the vents describe — `task create` reports success
+// (id + seq consumed) yet the task is never queryable. The real-world triggers
+// are FDA/sandbox losing fs access mid-write or another instance clobbering the
+// file; here we reproduce the observable symptom deterministically by writing an
+// empty array instead of the requested payload. Only the main tasks.json payload
+// write is intercepted; backups and every other fs op use the real impl.
+let dropTasksWrite = false;
+vi.mock("node:fs/promises", async () => {
+	const actual = await vi.importActual<typeof import("node:fs/promises")>("node:fs/promises");
+	// Matches the live tasks.json AND its `.tmp-<pid>` sibling (atomicWriteFile),
+	// but NOT the dated backup files under tasks-backups/.
+	const isTasksPayloadWrite = (p: string): boolean => !p.includes(".bak") && p.includes("tasks.json");
+	return {
+		...actual,
+		writeFile: vi.fn(async (path: any, data: any, opts?: any) => {
+			const p = String(path);
+			if (dropTasksWrite && isTasksPayloadWrite(p)) {
+				return actual.writeFile(p, "[]", opts);
+			}
+			return actual.writeFile(path, data, opts);
+		}),
+	};
+});
+
+const tempHome = mkdtempSync(join(tmpdir(), "dev3-create-verify-"));
+const dev3Home = join(tempHome, ".dev3.0");
+const originalHome = process.env.HOME;
+const tasksDir = join(dev3Home, "data", "tmp-existing-project");
+const tasksFile = join(tasksDir, "tasks.json");
+
+function makeProject(overrides?: Partial<Project>): Project {
+	return {
+		id: "proj-1",
+		name: "Existing Project",
+		path: "/tmp/existing-project",
+		setupScript: "",
+		setupScriptLaunchMode: "parallel",
+		devScript: "",
+		cleanupScript: "",
+		defaultBaseBranch: "main",
+		createdAt: "2026-04-15T00:00:00.000Z",
+		labels: [],
+		customColumns: [],
+		...overrides,
+	};
+}
+
+describe("addTask verifies persistence before returning", () => {
+	beforeEach(() => {
+		dropTasksWrite = false;
+		vi.resetModules();
+		process.env.HOME = tempHome;
+		rmSync(tempHome, { recursive: true, force: true });
+		mkdirSync(dev3Home, { recursive: true });
+		writeFileSync(join(dev3Home, "projects.json"), JSON.stringify([makeProject()], null, 2));
+	});
+
+	afterEach(() => {
+		dropTasksWrite = false;
+	});
+
+	afterAll(() => {
+		process.env.HOME = originalHome;
+		rmSync(tempHome, { recursive: true, force: true });
+	});
+
+	it("throws (instead of returning a ghost task) when the write is silently dropped", async () => {
+		const data = await import("../data");
+
+		dropTasksWrite = true;
+		await expect(data.addTask(makeProject(), "Ghost task", "todo")).rejects.toThrow(/persist/i);
+		dropTasksWrite = false;
+
+		// The task must genuinely NOT be on disk — confirming the throw reflects a
+		// real persistence failure, not a false alarm.
+		const onDisk = JSON.parse(readFileSync(tasksFile, "utf8")) as Task[];
+		expect(onDisk.find((t) => t.description === "Ghost task")).toBeUndefined();
+	});
+
+	it("returns a task that is immediately readable from disk on a healthy write", async () => {
+		const data = await import("../data");
+
+		const task = await data.addTask(makeProject(), "Real task", "todo");
+
+		const onDisk = JSON.parse(readFileSync(tasksFile, "utf8")) as Task[];
+		expect(onDisk.find((t) => t.id === task.id)).toBeDefined();
+		expect(task.description).toBe("Real task");
+	});
+});
diff --git a/src/bun/data.ts b/src/bun/data.ts
index 62044695..0fb18038 100644
--- a/src/bun/data.ts
+++ b/src/bun/data.ts
@@ -788,6 +788,24 @@ export async function addTask(
 		task.history = [{ at: now, title: getTaskTitle(task), overview: getTaskOverview(task), changed: "created" }];
 		tasks.push(task);
 		await rawSaveTasks(project, tasks);
+
+		// Verify the write actually landed before reporting success. atomicWriteFile
+		// can report success while the new content never reaches disk — e.g. macOS
+		// Full Disk Access / sandbox loss mid-write, or another running app instance
+		// clobbering the file. Without this guard the CLI prints "Created task <id>"
+		// (consuming a seq) for a task that is never queryable, which an agent then
+		// trusts. Re-read fresh from disk (strict bypasses the cache) and fail loudly
+		// instead of returning a ghost task. See decision 082.
+		const persisted = await rawLoadTasks(project, { strict: true });
+		if (!persisted.some((t) => t.id === task.id)) {
+			log.error("Task create verification failed — write did not persist", { taskId: task.id, seq: task.seq });
+			throw new Error(
+				`Task ${task.id} failed to persist (verification read-back did not find it). ` +
+				`The write reported success but the task is not on disk — likely macOS Full Disk Access loss ` +
+				`or another running app instance clobbering ${tasksFile(project)}.`,
+			);
+		}
+
 		log.info("Task created", { taskId: task.id, seq: task.seq, title });
 		return task;
 	});