feat(chat): add "Off" option to Reasoning Effort dropdown (thinking on/off)

Add an "Off" choice as the FIRST option in the per-conversation Reasoning
Effort selector. Selecting it disables extended thinking ENTIRELY for the
conversation's turns — NOT "set the effort to its lowest level". The two axes
are kept SEPARATE on the wire (per the umans API model, where thinking-off is
`reasoning_effort: "none"`, distinct from the effort level):

- `reasoningEffort` is the thinking-DEPTH ladder (low→max) — UNCHANGED.
- `thinking` is the on/off switch — a NEW per-conversation boolean.

Turning thinking off PRESERVES the persisted effort level, so an off→on toggle
restores the previously-chosen depth. The selector CONFLATES the two axes into
one <select> (UX); the wire does not.

Pure logic (reasoning-effort.ts): ThinkingSelection = "off" | ReasoningEffort,
selectionOptions() (Off first + the ladder, default marked), isThinkingSelection,
effectiveSelection(persistedEffort, persistedThinking), + FE-local PROPOSED wire
types (ThinkingResponse, SetThinkingRequest) + SaveThinkingSelection port. The
existing effortOptions()/isReasoningEffort()/effectiveEffort() are UNCHANGED — the
heartbeat feature reuses them (its own dropdown is unaffected).

Store: `thinking` state (boolean|null, null⇒ON default), refreshThinking() (GET,
alongside refreshReasoningEffort() on every focus/draft/switch), setThinking()
(PUT). App.svelte's saveThinkingSelection adapter fans one selection out: "off"
→ PUT /thinking {thinking:false} (effort untouched); a level → ensure thinking ON
(if off) then PUT /reasoning-effort {level}.

Backend contract gap (CR-14, backend-handoff.md §2l): the backend has NO
thinking-off mechanism today (umans mapReasoningEffort can't emit "none"). The FE
is built against the PROPOSED additive contract (GET/PUT /conversations/:id/thinking
+ ThinkingResponse/SetThinkingRequest; umans maps thinking===false →
reasoning_effort:"none"). Until shipped, GET /thinking 404s and refreshThinking()
leaves `thinking` null⇒ON — the selector gracefully shows the effort level, no crash.

Verification: typecheck 0/0, 1128 tests green (run TWICE — touches the shared fetch
fake), biome clean, build OK. 9 files (8 source + backend-handoff.md).

9 files changed, 549 insertions, 42 deletions

diff --git a/backend-handoff.md b/backend-handoff.md
index 4e3c6df..a9aaa5a 100644
--- a/backend-handoff.md
+++ b/backend-handoff.md
@@ -5,6 +5,17 @@
 > **From:** dispatch-web orchestrator · **To:** `../backend` orchestrator · **Courier:** the user.
 > `lsp` does NOT span the repos (AGENTS.md § Backend seam) — every cross-repo ask flows through here.
 
+_Last updated: 2026-06-28 (§2l — **thinking on/off** as a SEPARATE per-conversation axis from the
+reasoning-effort level. An `Off` option is now the FIRST choice in the Reasoning Effort dropdown;
+selecting it disables thinking entirely (the umans route maps it to `reasoning_effort: "none"`), NOT
+"effort to its lowest". The two axes are kept separate on the wire: a new `thinking: boolean|null`
+persisted per-conversation (null⇒ON default) + PROPOSED `GET`/`PUT /conversations/:id/thinking`
+(CR-14); the effort ladder + heartbeat dropdown are UNCHANGED. FE built against the proposed contract
+(FE-local `ThinkingResponse`/`SetThinkingRequest` in `reasoning-effort.ts`, swapped for the imported
+ones once the backend ships); a 404 (endpoint not yet shipped) leaves `thinking` null⇒ON, so the selector
+gracefully shows the effort level until then. typecheck 0/0, 1128 tests green (run TWICE), biome clean,
+build OK. Worktree env note: an untracked `dispatch-backend → backend` symlink in the worktree parent so
+the canonical `file:../dispatch-backend/...` paths resolve — NOT committed.)_
 _Last updated: 2026-06-27 (FE-only slice: **workspace-active indicator** — loading-dots on
 workspace cards when a workspace has ≥1 active/queued conversation. New `AppStore.workspaceHasActiveConversations(workspaceId)` derives from the existing open-tab set (every active/queued
 conversation has an open tab stamped with its `workspaceId`) × the backend lifecycle statuses; a
@@ -26,7 +37,7 @@ _Last updated: 2026-06-27 (workspace starring — backend `feature/workspace-sta
 return the updated `Workspace`). FE consumed: `adapter/http` `star`/`unstar`, pure `sortWorkspaces`/`applyStarred`,
 `store.setStarred` (optimistic + revert, `$derived`-sorted list), `WorkspaceCard` star toggle (filled gold ★ / outline ☆).
 Re-mirrored `.dispatch/wire.reference.md`. typecheck 0/0, 1045 tests green (+27), biome clean, build OK. No open backend asks.)_
-**FE is current on `[email protected]` / `[email protected]` / `[email protected]`.** Open asks: **CR-9**
+**FE is current on `[email protected]` / `[email protected]` / `[email protected]`.** Open asks: **CR-14** (thinking on/off — §2l; FE built against the proposed contract, awaiting the `GET`/`PUT /conversations/:id/thinking` endpoint + `ThinkingResponse`/`SetThinkingRequest` types), **CR-9**
 _Last updated: 2026-06-26 (backend: concurrency limits now PERSISTED across reboots — no API contract change, no FE
 re-pin/re-mirror needed; §2j updated. FE: brief "Saved." confirmation on the limit row after a successful save. 926 tests
 green.) Prior: CR-13 (`"queued"` ConversationStatus) RESOLVED; dev merged (e81df4c)._
@@ -1194,6 +1205,88 @@ click "Restore to N" → confirm the banner drops on the next poll.
 
 ---
 
+## 2l. Thinking on/off (separate from the reasoning-effort level) → **FE BUILT; 1 BACKEND ASK (CR-14)**
+
+A per-conversation **"thinking off"** affordance in the Reasoning Effort dropdown: an `Off` option as the
+FIRST choice (before `Low`). Selecting it disables extended thinking ENTIRELY for the conversation's turns
+— NOT "set the effort to its lowest level". The two axes are kept SEPARATE on the wire (per the umans API
+model): `reasoningEffort` is the thinking-DEPTH ladder (`low`→`max`); `thinking` is the on/off switch.
+Turning thinking off PRESERVES the persisted effort level, so an off→on toggle restores the previously-chosen
+depth. The per-conversation selector CONFLATES the two axes into one `<select>` (UX), but the WIRE does not.
+
+### Why a separate signal (not a new ladder level)
+
+The umans route (`provider-umans/src/reasoning.ts`) accepts `reasoning_effort: "none"|"low"|"medium"|"high"`
+— `"none"` disables thinking. The current `mapReasoningEffort` can only emit `"low"|"medium"|"high"` (it
+caps `xhigh`/`max`→`"high"`, and emits NO field when `reasoningEffort` is `undefined`, which the umans
+route treats as "default on"). So there is today NO path through the backend to `reasoning_effort: "none"`,
+i.e. no way to turn thinking off. Extending `ReasoningEffort` with `"none"` would work for umans but
+CONFLATES the two axes the user explicitly wants kept separate, and would also bleed into the heartbeat
+config dropdown (which reuses `effortOptions()`/`isReasoningEffort()`). A separate `thinking` boolean keeps
+the axes independent and leaves the existing ladder + heartbeat control untouched.
+
+### What the FE built (against the PROPOSED contract below)
+
+- Pure logic (`src/features/chat/reasoning-effort.ts`): a `ThinkingSelection = "off" | ReasoningEffort`
+  type, `selectionOptions()` (`Off` first, then the ladder w/ default marked), `isThinkingSelection()`,
+  `effectiveSelection(persistedEffort, persistedThinking)`, + FE-local proposed wire types
+  `ThinkingResponse` / `SetThinkingRequest` + a `SaveThinkingSelection` port. The existing
+  `effortOptions()` / `isReasoningEffort()` / `effectiveEffort()` are UNCHANGED (the heartbeat feature
+  reuses them — its own dropdown is unaffected).
+- Selector (`ReasoningEffortSelector.svelte`): renders `Off` first; new props `persistedEffort` +
+  `persistedThinking` + `save`; the displayed selection derives from BOTH axes (`thinking===false` ⇒ `off`,
+  else the effective effort level). Selecting `Off` calls `save("off")`; selecting a level calls
+  `save(level)`.
+- Store (`store.svelte.ts`): `thinking` state (`boolean | null`, `null`=never set⇒ON), `refreshThinking()`
+  (GET, called alongside `refreshReasoningEffort()` on every focus/draft/switch), `setThinking()`
+  (PUT). `App.svelte`'s `saveThinkingSelection` adapter fans one selection out to the right PUT(s):
+  `"off"` → `PUT /thinking {thinking:false}` (leaves effort untouched); a level → ensure thinking ON
+  (`PUT /thinking {thinking:true}` if currently off) then `PUT /reasoning-effort {level}`.
+- **Graceful pending-backend behavior:** `refreshThinking()` returns on `!res.ok` (a 404 — endpoint not
+  yet shipped — leaves `thinking` `null` ⇒ ON/default), so the selector simply shows the effort level
+  until the backend ships CR-14. No crash, no broken UI.
+- Tests: pure-logic (`selectionOptions`/`isThinkingSelection`/`effectiveSelection` — incl. `thinking===false`
+  ⇒ `off` while a level is still persisted), component (`Off` first, selecting `Off` sends `"off"` not a
+  level, off shows `off` even with a persisted level, failed-save revert, in-flight disable), store
+  (seeds from GET, treats a 404 as null⇒ON, PUT round-trips + echoes). typecheck 0/0, **1128 tests green**
+  (run TWICE — touches the shared fetch fake), biome clean, build OK.
+
+### CR-14 — PROPOSED backend contract (the ask)
+
+A NEW, ADDITIVE per-conversation "thinking" boolean, fully separate from `reasoningEffort`. Mirrors the
+existing `reasoning-effort` shape so it slots in with minimal surface:
+
+- **Wire / transport-contract (additive, NO version bump needed):**
+  ```ts
+  // Response of GET /conversations/:id/thinking
+  export interface ThinkingResponse {
+    readonly conversationId: string;
+    readonly thinking: boolean | null;   // null = never set ⇒ thinking ON (the default)
+  }
+  // Body of PUT /conversations/:id/thinking
+  export interface SetThinkingRequest { readonly thinking: boolean }
+  ```
+  Optionally also a per-turn `ChatRequest.thinking?: boolean` override (mirrors `reasoningEffort`).
+- **Endpoints (new, mirror `reasoning-effort`):**
+  - `GET /conversations/:id/thinking` → `ThinkingResponse`
+  - `PUT /conversations/:id/thinking` (body `SetThinkingRequest`) → `ThinkingResponse`
+- **Resolution (server-owned — do NOT re-implement in the FE):** per-turn `thinking` → persisted
+  conversation `thinking` → default **ON** (`true`). When resolved **OFF**, the orchestrator signals the
+  provider to DISABLE extended thinking; `provider-umans` maps that to `reasoning_effort: "none"` (its
+  `mapReasoningEffort` would gain a `thinking===false ⇒ "none"` short-circuit, ignoring the effort level
+  while off). Providers without a thinking knob ignore it (safe, as today).
+- **No change to `ReasoningEffort` / `reasoning-effort` endpoint** — the ladder and its validation are
+  untouched (so the heartbeat config + existing callers are unaffected).
+
+**FE re-pin/re-mirror on shipment:** once the backend ships CR-14, move `ThinkingResponse` /
+`SetThinkingRequest` into `@dispatch/transport-contract`, re-pin the `file:` dep, re-mirror
+`.dispatch/transport-contract.reference.md`, and swap the FE-local types for the imported ones (the store
++ `reasoning-effort.ts` already reference them by name — a one-line import-source change). The graceful-404
+fallback can then be tightened. **A live integration probe is pending** the endpoint existing
+(`scripts/live-probe.ts` would gain a `/thinking` round-trip).
+
+---
+
 ## 3. Likely NEXT backend asks (heads-up, not yet requested)
 
 - **Model max context-window LIMIT** → **CONSUMED ✅** — `GET /models` now returns
diff --git a/src/app/App.svelte b/src/app/App.svelte
index 7e5ac7d..f87eddb 100644
--- a/src/app/App.svelte
+++ b/src/app/App.svelte
@@ -1,5 +1,5 @@
 <script lang="ts">
-	import type { ImageInput, ReasoningEffort } from "@dispatch/transport-contract";
+	import type { ImageInput } from "@dispatch/transport-contract";
 	import type { InvokeMessage } from "@dispatch/ui-contract";
 	import { tick } from "svelte";
 	import Table from "../components/Table.svelte";
@@ -17,8 +17,9 @@
 		ReasoningEffortSelector,
 		type CompactNowResult,
 		type ComposerStatus,
-		type ReasoningEffortSaveResult,
 		type SaveCompactPercentResult,
+		type ThinkingSelection,
+		type ThinkingSelectionSaveResult,
 	} from "../features/chat";
 	import { manifest as conversationCacheManifest } from "../features/conversation-cache";
 	import { manifest as markdownManifest } from "../features/markdown";
@@ -317,14 +318,35 @@
 			: { ok: false, error: result.error };
 	}
 
-	// Adapt the store's reasoning-effort result to the chat feature's port.
-	async function saveReasoningEffort(
-		level: ReasoningEffort,
-	): Promise<ReasoningEffortSaveResult | null> {
-		const result = await store.setReasoningEffort(level);
+	// Adapt the store's reasoning-effort + thinking results to the chat
+	// feature's combined selector port. The selector sends ONE selection ("off"
+	// or a level); the adapter fans it out to the right per-axis PUT(s). "off" is
+	// a SEPARATE signal from the effort level: it persists `thinking: false`
+	// (the umans route maps that to `reasoning_effort: "none"`), leaving the
+	// effort level untouched so an off→on toggle restores it. A level ensures
+	// thinking is ON (the level is meaningless while thinking is off) then sets
+	// the effort level.
+	async function saveThinkingSelection(
+		selection: ThinkingSelection,
+	): Promise<ThinkingSelectionSaveResult | null> {
+		if (selection === "off") {
+			const result = await store.setThinking(false);
+			if (result === null) return null;
+			return result.ok
+				? { ok: true, selection: "off" }
+				: { ok: false, error: result.error };
+		}
+		// A level: enable thinking first if it is currently off, then set the level.
+		if (store.thinking === false) {
+			const on = await store.setThinking(true);
+			if (on !== null && !on.ok) {
+				return { ok: false, error: on.error };
+			}
+		}
+		const result = await store.setReasoningEffort(selection);
 		if (result === null) return null;
 		return result.ok
-			? { ok: true, reasoningEffort: result.reasoningEffort }
+			? { ok: true, selection: result.reasoningEffort }
 			: { ok: false, error: result.error };
 	}
 
@@ -728,7 +750,11 @@
 			     re-mount per conversation — incl. switching between drafts — and can't
 			     bleed across tabs. Editable for a draft too (cwd + effort apply from turn 1). -->
 			{#key store.currentConversationId}
-				<ReasoningEffortSelector persisted={store.reasoningEffort} save={saveReasoningEffort} />
+				<ReasoningEffortSelector
+					persistedEffort={store.reasoningEffort}
+					persistedThinking={store.thinking}
+					save={saveThinkingSelection}
+				/>
 				<CwdField cwd={store.cwd} canEdit={true} save={saveCwd} />
 				<ComputerField
 					computerId={store.computerId}
diff --git a/src/app/store.svelte.ts b/src/app/store.svelte.ts
index 22b0a25..e0491ab 100644
--- a/src/app/store.svelte.ts
+++ b/src/app/store.svelte.ts
@@ -54,6 +54,7 @@ import {
 } from "../core/protocol";
 import type { ChatStore, HistorySync, MetricsSync } from "../features/chat";
 import { createChatStore } from "../features/chat";
+import type { SetThinkingRequest, ThinkingResponse } from "../features/chat/reasoning-effort";
 import type {
   ConcurrencyCooldownResult,
   ConcurrencyDeleteResult,
@@ -132,6 +133,14 @@ export type ReasoningEffortResult =
   | { readonly ok: true; readonly reasoningEffort: ReasoningEffort }
   | { readonly ok: false; readonly error: string };
 
+/**
+ * Outcome of `PUT /conversations/:id/thinking` (PROPOSED — see
+ * `backend-handoff.md`; the endpoint is not yet shipped by the backend).
+ */
+export type ThinkingResult =
+  | { readonly ok: true; readonly thinking: boolean }
+  | { readonly ok: false; readonly error: string };
+
 /** Outcome of `POST /conversations/:id/compact` (manual compaction). */
 export type CompactResult =
   | { readonly ok: true; readonly response: CompactResponse }
@@ -273,6 +282,22 @@ export interface AppStore {
    */
   setReasoningEffort(level: ReasoningEffort): Promise<ReasoningEffortResult | null>;
   /**
+   * The workspace conversation's persisted thinking flag, or null when never
+   * set (the server then resolves turns with thinking ON — the default).
+   * `false` ⇒ thinking disabled entirely (the SEPARATE "off" axis — NOT a
+   * zero-effort level; the umans route maps it to `reasoning_effort: "none"`).
+   * PROPOSED backend contract — see `backend-handoff.md`.
+   */
+  readonly thinking: boolean | null;
+  /**
+   * Persist the workspace conversation's thinking flag
+   * (`PUT /conversations/:id/thinking`). Works for a draft too (its id survives
+   * promotion), so the first turn already runs with the chosen setting. Takes
+   * effect from the NEXT turn; resolution stays server-owned.
+   * PROPOSED backend contract — see `backend-handoff.md`.
+   */
+  setThinking(enabled: boolean): Promise<ThinkingResult | null>;
+  /**
    * Manually trigger conversation compaction (`POST /conversations/:id/compact`).
    * Summarizes old messages + retains the most recent N. Returns null when no
    * conversation is focused (a draft has nothing to compact).
@@ -729,6 +754,30 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
     }
   }
 
+  // The workspace conversation's persisted thinking flag (SEPARATE from the
+  // effort level). Seeded from the backend on focus change; null = never set
+  // (thinking ON — the default). PROPOSED endpoint (see backend-handoff.md):
+  // a 404 (endpoint not yet shipped) leaves `thinking` null ⇒ ON (default), so
+  // the selector simply shows the effort level until the backend ships it.
+  let thinking = $state<boolean | null>(null);
+
+  /** Refetch the workspace conversation's thinking flag (works for a draft too). */
+  async function refreshThinking(): Promise<void> {
+    const id = workspaceConversationId();
+    // Clear immediately so a switch never shows the PREVIOUS conversation's
+    // setting while the fetch is in flight (null ⇒ ON, the default).
+    thinking = null;
+    try {
+      const res = await fetchImpl(`${httpBase}/conversations/${encodeURIComponent(id)}/thinking`);
+      if (!res.ok) return;
+      const data = (await res.json()) as ThinkingResponse;
+      // Guard a slow response losing a race with a conversation switch.
+      if (workspaceConversationId() === id) thinking = data.thinking ?? null;
+    } catch (err) {
+      reportError("Failed to load thinking setting", err);
+    }
+  }
+
   // The workspace conversation's auto-compact percent. Seeded from the
   // backend on focus change; null = not yet fetched. 0 = disabled.
   let compactPercent = $state<number | null>(null);
@@ -975,6 +1024,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
     void refreshCwd();
     void refreshComputer();
     void refreshReasoningEffort();
+    void refreshThinking();
     void refreshCompactPercent();
   }
 
@@ -1208,6 +1258,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
   void refreshComputer();
   void refreshModel();
   void refreshReasoningEffort();
+  void refreshThinking();
   void refreshCompactPercent();
   void refreshVisionSettings();
 
@@ -1243,6 +1294,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
       void refreshComputer();
       void refreshModel();
       void refreshReasoningEffort();
+      void refreshThinking();
       void refreshCompactPercent();
     },
     get activeChat(): ChatStore {
@@ -1286,6 +1338,9 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
     get reasoningEffort(): ReasoningEffort | null {
       return reasoningEffort;
     },
+    get thinking(): boolean | null {
+      return thinking;
+    },
     get compactPercent(): number | null {
       return compactPercent;
     },
@@ -1340,6 +1395,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
         void refreshCwd();
         void refreshComputer();
         void refreshReasoningEffort();
+        void refreshThinking();
         void refreshCompactPercent();
         // Now send on the promoted store
         chatStores.get(conversationId)?.send(text, images);
@@ -1386,6 +1442,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
       void refreshComputer();
       void refreshModel();
       void refreshReasoningEffort();
+      void refreshThinking();
       void refreshCompactPercent();
     },
 
@@ -1401,6 +1458,7 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
       void refreshComputer();
       void refreshModel();
       void refreshReasoningEffort();
+      void refreshThinking();
       void refreshCompactPercent();
     },
 
@@ -1587,6 +1645,37 @@ export function createAppStore(opts?: CreateAppStoreOptions): AppStore {
       }
     },
 
+    async setThinking(enabled: boolean): Promise<ThinkingResult | null> {
+      const id = workspaceConversationId();
+      const body: SetThinkingRequest = { thinking: enabled };
+      try {
+        const res = await fetchImpl(
+          `${httpBase}/conversations/${encodeURIComponent(id)}/thinking`,
+          {
+            method: "PUT",
+            headers: { "content-type": "application/json" },
+            body: JSON.stringify(body),
+          },
+        );
+        if (!res.ok) {
+          const errBody = (await res.json().catch(() => null)) as { error?: string } | null;
+          return {
+            ok: false,
+            error: errBody?.error ?? `Set thinking failed (HTTP ${res.status})`,
+          };
+        }
+        const data = (await res.json()) as ThinkingResponse;
+        const next = data.thinking ?? enabled;
+        if (workspaceConversationId() === id) thinking = next;
+        return { ok: true, thinking: next };
+      } catch (err) {
+        return {
+          ok: false,
+          error: err instanceof Error ? err.message : "Set thinking request failed",
+        };
+      }
+    },
+
     stopGeneration(): void {
       const conversationId = tabsStore.activeConversationId;
       if (conversationId === null) return;
diff --git a/src/app/store.test.ts b/src/app/store.test.ts
index a945ea7..dbceec5 100644
--- a/src/app/store.test.ts
+++ b/src/app/store.test.ts
@@ -891,6 +891,100 @@ describe("createAppStore", () => {
     store.dispose();
   });
 
+  it("seeds thinking from GET /conversations/:id/thinking (null = never set ⇒ ON)", async () => {
+    const base = fakeFetchImpl();
+    const fetchImpl: typeof fetch = async (input, init) => {
+      const url = typeof input === "string" ? input : input instanceof URL ? input.href : input.url;
+      if (url.endsWith("/thinking")) {
+        return new Response(JSON.stringify({ conversationId: "x", thinking: false }), {
+          status: 200,
+        });
+      }
+      return base(input, init);
+    };
+    const ws = fakeSocket();
+    const store = createAppStore({
+      socketFactory: () => ws,
+      fetchImpl,
+      localStorage: createFakeStorage(),
+    });
+    ws.resolveOpen();
+
+    await vi.waitFor(() => {
+      expect(store.thinking).toBe(false);
+    });
+
+    store.dispose();
+  });
+
+  it("treats a missing thinking endpoint (404) as 'never set' (null ⇒ ON)", async () => {
+    // The thinking endpoint is PROPOSED (backend-handoff.md); until the backend
+    // ships it, GET 404s and `thinking` stays null ⇒ the selector shows the
+    // effort level (thinking ON, the default) — graceful, never a crash.
+    const base = fakeFetchImpl();
+    const fetchImpl: typeof fetch = async (input, init) => {
+      const url = typeof input === "string" ? input : input instanceof URL ? input.href : input.url;
+      if (url.endsWith("/thinking")) {
+        return new Response("not found", { status: 404 });
+      }
+      return base(input, init);
+    };
+    const ws = fakeSocket();
+    const store = createAppStore({
+      socketFactory: () => ws,
+      fetchImpl,
+      localStorage: createFakeStorage(),
+    });
+    ws.resolveOpen();
+
+    // give the (404ing) fetch a tick to settle
+    await vi.waitFor(() => {
+      expect(store.reasoningEffort).not.toBe(undefined);
+    });
+    expect(store.thinking).toBeNull();
+
+    store.dispose();
+  });
+
+  it("setThinking PUTs the flag and updates local state from the echo", async () => {
+    const calls: { url: string; method: string; body: string | undefined }[] = [];
+    const base = fakeFetchImpl();
+    const fetchImpl: typeof fetch = async (input, init) => {
+      const url = typeof input === "string" ? input : input instanceof URL ? input.href : input.url;
+      calls.push({ url, method: init?.method ?? "GET", body: init?.body as string | undefined });
+      if (url.endsWith("/thinking") && init?.method === "PUT") {
+        const sent = JSON.parse(init.body as string) as { thinking: boolean };
+        return new Response(JSON.stringify({ conversationId: "x", thinking: sent.thinking }), {
+          status: 200,
+        });
+      }
+      if (url.endsWith("/thinking")) {
+        return new Response(JSON.stringify({ conversationId: "x", thinking: null }), {
+          status: 200,
+        });
+      }
+      return base(input, init);
+    };
+    const ws = fakeSocket();
+    const store = createAppStore({
+      socketFactory: () => ws,
+      fetchImpl,
+      localStorage: createFakeStorage(),
+    });
+    ws.resolveOpen();
+
+    const result = await store.setThinking(false);
+    expect(result).toEqual({ ok: true, thinking: false });
+    expect(store.thinking).toBe(false);
+
+    const put = calls.find((c) => c.method === "PUT" && c.url.endsWith("/thinking"));
+    expect(put).toBeDefined();
+    expect(put?.url).toContain(`/conversations/${store.currentConversationId}/`);
+    expect(JSON.parse(put?.body ?? "{}")).toEqual({ thinking: false });
+
+    store.dispose();
+  });
+
   it("does NOT re-scope a scope:'global' surface on conversation switch (no churn)", () => {
     const ws = fakeSocket();
     const store = createAppStore({
diff --git a/src/features/chat/index.ts b/src/features/chat/index.ts
index cf57cea..2f98a0e 100644
--- a/src/features/chat/index.ts
+++ b/src/features/chat/index.ts
@@ -12,13 +12,23 @@ export type {
   EffortOption,
   ReasoningEffortSaveResult,
   SaveReasoningEffort,
+  SaveThinkingSelection,
+  SelectionOption,
+  SetThinkingRequest,
+  ThinkingResponse,
+  ThinkingSaveResult,
+  ThinkingSelection,
+  ThinkingSelectionSaveResult,
 } from "./reasoning-effort";
 export {
   DEFAULT_REASONING_EFFORT,
   effectiveEffort,
+  effectiveSelection,
   effortOptions,
   isReasoningEffort,
+  isThinkingSelection,
   REASONING_EFFORT_LEVELS,
+  selectionOptions,
 } from "./reasoning-effort";
 export type { ChatStore, ChatStoreDependencies } from "./store.svelte";
 export { createChatStore } from "./store.svelte";
diff --git a/src/features/chat/reasoning-effort.test.ts b/src/features/chat/reasoning-effort.test.ts
index 6d409e9..e870bac 100644
--- a/src/features/chat/reasoning-effort.test.ts
+++ b/src/features/chat/reasoning-effort.test.ts
@@ -2,9 +2,12 @@ import { describe, expect, it } from "vitest";
 import {
   DEFAULT_REASONING_EFFORT,
   effectiveEffort,
+  effectiveSelection,
   effortOptions,
   isReasoningEffort,
+  isThinkingSelection,
   REASONING_EFFORT_LEVELS,
+  selectionOptions,
 } from "./reasoning-effort";
 
 describe("reasoning-effort helpers", () => {
@@ -43,3 +46,39 @@ describe("reasoning-effort helpers", () => {
     }
   });
 });
+
+describe("thinking selection (the separate on/off axis)", () => {
+  it("selectionOptions lists 'off' first, then the ladder (default marked)", () => {
+    const options = selectionOptions();
+    expect(options).toHaveLength(1 + REASONING_EFFORT_LEVELS.length);
+    expect(options[0]?.value).toBe("off");
+    expect(options[0]?.label).toBe("Off");
+    // the rest are the ladder, unchanged from effortOptions()
+    expect(options.slice(1).map((o) => o.value)).toEqual([...REASONING_EFFORT_LEVELS]);
+    expect(options.find((o) => o.value === "high")?.label).toBe("high (default)");
+  });
+
+  it("isThinkingSelection narrows 'off' + ladder strings, rejects the rest", () => {
+    expect(isThinkingSelection("off")).toBe(true);
+    for (const level of REASONING_EFFORT_LEVELS) {
+      expect(isThinkingSelection(level)).toBe(true);
+    }
+    expect(isThinkingSelection("banana")).toBe(false);
+    expect(isThinkingSelection("")).toBe(false);
+    expect(isThinkingSelection("OFF")).toBe(false);
+    expect(isThinkingSelection("none")).toBe(false); // NOT a wire value we send
+  });
+
+  it("effectiveSelection shows 'off' when thinking is explicitly disabled", () => {
+    // thinking off is a SEPARATE axis: the effort level is irrelevant while off.
+    expect(effectiveSelection("xhigh", false)).toBe("off");
+    expect(effectiveSelection(null, false)).toBe("off");
+  });
+
+  it("effectiveSelection shows the effort level when thinking is on (default)", () => {
+    // null thinking = never set ⇒ thinking ON (default) ⇒ show the effort level.
+    expect(effectiveSelection(null, null)).toBe("high"); // default effort
+    expect(effectiveSelection("low", null)).toBe("low");
+    expect(effectiveSelection("max", true)).toBe("max"); // explicitly on
+  });
+});
diff --git a/src/features/chat/reasoning-effort.ts b/src/features/chat/reasoning-effort.ts
index 1eb77b6..39e1c5a 100644
--- a/src/features/chat/reasoning-effort.ts
+++ b/src/features/chat/reasoning-effort.ts
@@ -36,7 +36,7 @@ export function effectiveEffort(persisted: ReasoningEffort | null): ReasoningEff
   return persisted ?? DEFAULT_REASONING_EFFORT;
 }
 
-/** One `<option>` of the selector. */
+/** One `<option>` of the effort ladder. */
 export interface EffortOption {
   readonly value: ReasoningEffort;
   readonly label: string;
@@ -64,3 +64,99 @@ export type ReasoningEffortSaveResult =
 export type SaveReasoningEffort = (
   level: ReasoningEffort,
 ) => Promise<ReasoningEffortSaveResult | null>;
+
+// ── Thinking on/off (a SEPARATE axis from the effort level) ─────────────────
+//
+// Per the umans API (and the user's mental model), "thinking off" is NOT a
+// zero-effort level — it is a distinct "disable extended thinking entirely"
+// signal. The umans route expresses it as `reasoning_effort: "none"`; Dispatch
+// surfaces it as a SEPARATE per-conversation boolean so the effort LEVEL is
+// preserved across an off→on toggle (turning thinking back on restores the
+// previously-chosen depth). The per-conversation selector CONFLATES the two
+// axes into one `<select>` (UX), but the WIRE keeps them separate.
+//
+// ⚠️ BACKEND CONTRACT GAP — see `backend-handoff.md`. The `thinking` endpoint +
+// wire types below are the PROPOSED shape; the backend has NOT shipped them yet.
+// They are defined FE-local (mirroring the shipped `ReasoningEffortResponse` /
+// `SetReasoningEffortRequest` shape) so the FE is built + tested against the
+// target contract. Re-pin + re-mirror once the backend ships them.
+
+/**
+ * Response of `GET /conversations/:id/thinking` (PROPOSED). `thinking` is null
+ * when never set (the server then resolves turns with thinking ON — the
+ * default), `false` when explicitly disabled, `true` when explicitly enabled.
+ */
+export interface ThinkingResponse {
+  readonly conversationId: string;
+  readonly thinking: boolean | null;
+}
+
+/** Body of `PUT /conversations/:id/thinking` (PROPOSED). */
+export interface SetThinkingRequest {
+  readonly thinking: boolean;
+}
+
+/**
+ * The per-conversation selector's value: `"off"` (thinking disabled — the
+ * separate axis) or a reasoning-effort LEVEL. NOT a widened ladder: `"off"` is
+ * not a degree of effort, it is the absence of thinking.
+ */
+export type ThinkingSelection = "off" | ReasoningEffort;
+
+/** One `<option>` of the combined selector (off or a level). */
+export interface SelectionOption {
+  readonly value: ThinkingSelection;
+  readonly label: string;
+}
+
+/**
+ * The selector's options: `"off"` first (the separate disable signal), then the
+ * effort ladder with the server default marked `(default)`. A never-set
+ * conversation (thinking on, effort null) reads "high (default)".
+ */
+export function selectionOptions(): readonly SelectionOption[] {
+  return [{ value: "off", label: "Off" }, ...effortOptions()];
+}
+
+/** Narrow an untrusted `<select>` value to a {@link ThinkingSelection}. */
+export function isThinkingSelection(value: string): value is ThinkingSelection {
+  return value === "off" || isReasoningEffort(value);
+}
+
+/**
+ * The selection the per-conversation selector should show as selected: `"off"`
+ * when thinking is explicitly disabled (`persistedThinking === false`), else the
+ * effective effort level. `persistedThinking === null` (never set) ⇒ thinking
+ * ON (the default) ⇒ the effort level is shown — NOT "off".
+ */
+export function effectiveSelection(
+  persistedEffort: ReasoningEffort | null,
+  persistedThinking: boolean | null,
+): ThinkingSelection {
+  if (persistedThinking === false) return "off";
+  return effectiveEffort(persistedEffort);
+}
+
+// ── Injected port for the combined selector (off OR a level) ─────────────────
+
+/** Outcome of persisting a {@link ThinkingSelection} (one or two PUTs). */
+export type ThinkingSelectionSaveResult =
+  | { readonly ok: true; readonly selection: ThinkingSelection }
+  | { readonly ok: false; readonly error: string };
+
+/**
+ * Persist a thinking selection (consumer-defines-port; the composition root
+ * adapts the store's `PUT .../thinking` + `PUT .../reasoning-effort` to this).
+ * - `"off"` → disable thinking (the separate signal); the effort level is left
+ *   untouched so an off→on toggle restores it.
+ * - a level → set the effort level AND ensure thinking is ON (the level is
+ *   meaningless while thinking is off).
+ */
+export type SaveThinkingSelection = (
+  selection: ThinkingSelection,
+) => Promise<ThinkingSelectionSaveResult | null>;
+
+/** Outcome of `PUT /conversations/:id/thinking`. */
+export type ThinkingSaveResult =
+  | { readonly ok: true; readonly thinking: boolean }
+  | { readonly ok: false; readonly error: string };
diff --git a/src/features/chat/ui.test.ts b/src/features/chat/ui.test.ts
index 5f8067d..f4006f7 100644
--- a/src/features/chat/ui.test.ts
+++ b/src/features/chat/ui.test.ts
@@ -1092,30 +1092,53 @@ describe("ModelSelector", () => {
 });
 
 describe("ReasoningEffortSelector", () => {
-  it("renders null (never set) as the default level, marked '(default)'", () => {
-    render(ReasoningEffortSelector, { props: { persisted: null, save: vi.fn() } });
+  it("renders null effort + null thinking (never set) as the default level, marked '(default)'", () => {
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: null, persistedThinking: null, save: vi.fn() },
+    });
 
     const select = screen.getByRole("combobox", { name: "Reasoning effort" });
     expect(select).toHaveValue("high");
     expect(within(select).getByRole("option", { name: "high (default)" })).toBeInTheDocument();
-    // All five ladder levels are offered.
-    expect(within(select).getAllByRole("option")).toHaveLength(5);
+    // "Off" first, then the five ladder levels.
+    const options = within(select).getAllByRole("option");
+    expect(options).toHaveLength(6);
+    expect(options[0]).toHaveValue("off");
+    expect(within(select).getByRole("option", { name: "Off" })).toBeInTheDocument();
   });
 
-  it("renders a persisted level as selected", () => {
-    render(ReasoningEffortSelector, { props: { persisted: "xhigh", save: vi.fn() } });
+  it("renders a persisted level as selected when thinking is on", () => {
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: "xhigh", persistedThinking: null, save: vi.fn() },
+    });
 
     expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toHaveValue("xhigh");
   });
 
+  it("renders 'off' as selected when thinking is disabled (effort level is preserved but hidden)", () => {
+    // thinking off is a SEPARATE axis: even with a persisted effort level, the
+    // selector shows "off" while thinking is disabled.
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: "xhigh", persistedThinking: false, save: vi.fn() },
+    });
+
+    expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toHaveValue("off");
+    // the level option is still present (restored on an off→on toggle)
+    expect(
+      within(screen.getByRole("combobox")).getByRole("option", { name: "xhigh" }),
+    ).toBeInTheDocument();
+  });
+
   it("selecting a level saves it via the injected port and confirms", async () => {
-    const save = vi.fn(async (level: "low" | "medium" | "high" | "xhigh" | "max") => ({
+    const save = vi.fn(async (selection: "off" | "low" | "medium" | "high" | "xhigh" | "max") => ({
       ok: true as const,
-      reasoningEffort: level,
+      selection,
     }));
     const user = userEvent.setup();
 
-    render(ReasoningEffortSelector, { props: { persisted: null, save } });
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: null, persistedThinking: null, save },
+    });
 
     await user.selectOptions(screen.getByRole("combobox", { name: "Reasoning effort" }), "max");
 
@@ -1127,11 +1150,35 @@ describe("ReasoningEffortSelector", () => {
     expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toHaveValue("max");
   });
 
-  it("a failed save shows the error and reverts to the persisted value", async () => {
+  it("selecting 'off' saves the separate disable signal (not a level)", async () => {
+    const save = vi.fn(async (selection: "off" | "low" | "medium" | "high" | "xhigh" | "max") => ({
+      ok: true as const,
+      selection,
+    }));
+    const user = userEvent.setup();
+
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: "high", persistedThinking: null, save },
+    });
+
+    await user.selectOptions(screen.getByRole("combobox", { name: "Reasoning effort" }), "off");
+
+    expect(save).toHaveBeenCalledTimes(1);
+    // "off" is the separate thinking-disable signal — NOT a zero-effort level.
+    expect(save).toHaveBeenCalledWith("off");
+    await vi.waitFor(() => {
+      expect(screen.getByText(/applies from the next turn/i)).toBeInTheDocument();
+    });
+    expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toHaveValue("off");
+  });
+
+  it("a failed save shows the error and reverts to the persisted selection", async () => {
     const save = vi.fn(async () => ({ ok: false as const, error: "nope" }));
     const user = userEvent.setup();
 
-    render(ReasoningEffortSelector, { props: { persisted: "low", save } });
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: "low", persistedThinking: null, save },
+    });
 
     await user.selectOptions(screen.getByRole("combobox", { name: "Reasoning effort" }), "max");
 
@@ -1142,22 +1189,24 @@ describe("ReasoningEffortSelector", () => {
   });
 
   it("disables the select while a save is in flight (no double-fire)", async () => {
-    let resolveSave: ((r: { ok: true; reasoningEffort: "max" }) => void) | undefined;
+    let resolveSave: ((r: { ok: true; selection: "max" }) => void) | undefined;
     const save = vi.fn(
       () =>
-        new Promise<{ ok: true; reasoningEffort: "max" }>((resolve) => {
+        new Promise<{ ok: true; selection: "max" }>((resolve) => {
           resolveSave = resolve;
         }),
     );
     const user = userEvent.setup();
 
-    render(ReasoningEffortSelector, { props: { persisted: null, save } });
+    render(ReasoningEffortSelector, {
+      props: { persistedEffort: null, persistedThinking: null, save },
+    });
 
     await user.selectOptions(screen.getByRole("combobox", { name: "Reasoning effort" }), "max");
 
     expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toBeDisabled();
 
-    resolveSave?.({ ok: true, reasoningEffort: "max" });
+    resolveSave?.({ ok: true, selection: "max" });
     await vi.waitFor(() => {
       expect(screen.getByRole("combobox", { name: "Reasoning effort" })).toBeEnabled();
     });
diff --git a/src/features/chat/ui/ReasoningEffortSelector.svelte b/src/features/chat/ui/ReasoningEffortSelector.svelte
index d982905..6858779 100644
--- a/src/features/chat/ui/ReasoningEffortSelector.svelte
+++ b/src/features/chat/ui/ReasoningEffortSelector.svelte
@@ -1,34 +1,45 @@
 <script lang="ts">
   import type { ReasoningEffort } from "@dispatch/transport-contract";
   import {
-    effectiveEffort,
-    effortOptions,
-    isReasoningEffort,
-    type SaveReasoningEffort,
+    effectiveSelection,
+    isThinkingSelection,
+    selectionOptions,
+    type SaveThinkingSelection,
+    type ThinkingSelection,
   } from "../reasoning-effort";
 
   let {
-    persisted,
+    persistedEffort,
+    persistedThinking,
     save,
   }: {
-    /** The conversation's persisted level, or null when never set (default applies). */
-    persisted: ReasoningEffort | null;
-    save: SaveReasoningEffort;
+    /** The conversation's persisted effort level, or null when never set (default applies). */
+    persistedEffort: ReasoningEffort | null;
+    /**
+     * The conversation's persisted thinking flag, or null when never set
+     * (thinking ON — the default). `false` ⇒ thinking disabled (the separate
+     * "off" axis); the effort level is preserved across an off→on toggle.
+     */
+    persistedThinking: boolean | null;
+    /** Persist a thinking selection (off or a level). */
+    save: SaveThinkingSelection;
   } = $props();
 
-  const options = effortOptions();
+  const options = selectionOptions();
 
-  // The user's in-flight choice; null = mirror the (async-loaded) persisted prop.
-  // Re-mounted per conversation, so there is no cross-tab bleed.
-  let chosen = $state<ReasoningEffort | null>(null);
+  // The user's in-flight choice; null = mirror the (async-loaded) persisted
+  // selection. Re-mounted per conversation, so there is no cross-tab bleed.
+  let chosen = $state<ThinkingSelection | null>(null);
   let saving = $state(false);
   let error = $state<string | null>(null);
   let justSaved = $state(false);
 
-  const selected = $derived(chosen ?? effectiveEffort(persisted));
+  const selected = $derived(
+    chosen ?? effectiveSelection(persistedEffort, persistedThinking),
+  );
 
   async function handleChange(value: string) {
-    if (!isReasoningEffort(value) || saving) return;
+    if (!isThinkingSelection(value) || saving) return;
     chosen = value;
     saving = true;
     error = null;
@@ -40,7 +51,7 @@
       justSaved = true;
     } else {
       error = result.error;
-      chosen = null; // revert to the persisted value
+      chosen = null; // revert to the persisted selection
     }
   }
 </script>
@@ -69,7 +80,7 @@
     <p class="text-xs text-success">Saved — applies from the next turn.</p>
   {:else}
     <p class="text-xs opacity-50">
-      How long the model thinks before answering. Changing it can re-prefill the prompt cache once.
+      How long the model thinks before answering. “Off” disables thinking entirely. Changing it can re-prefill the prompt cache once.
     </p>
   {/if}
 </div>