From cac8e168bbc56a15140b7f9b51c4c8403304fb17 Mon Sep 17 00:00:00 2001
From: zszz3 <91608029+zszz3@users.noreply.github.com>
Date: Sun, 18 Oct 2026 00:40:30 +0800
Subject: [PATCH 1/3] feat(session): add /fast for OpenAI's priority service
 tier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

ChatGPT subscribers and OpenAI API users can ask for a faster priority
tier (Responses `service_tier: "priority"`), which costs more plan
limits or a higher per-token price. PI-Desktop never requested a tier.

Model settings now own whether a model may use it: a per-model "Fast
mode" opt-in, enabled for Responses wires, stored in the existing
binding JSON. The cost is a per-conversation choice, so `/fast` toggles
a persisted session switch (schema v25, additive). On a model that is
not configured, `/fast` stores nothing and points to Settings → Models →
Advanced. Launch sends Fast mode only when both are on, and only main
turns carry the field; subagents and compaction keep the default tier.

closes #1657
---
 .../desktop/electron/main/builtin-commands.ts |   1 +
 apps/desktop/electron/main/ipc/session-ipc.ts |   1 +
 .../electron/main/remote/remote-backend.ts    |   9 +-
 .../electron/main/runtime/session-launch.ts   |   3 +
 apps/desktop/src/components/Composer.tsx      |   2 +
 apps/desktop/src/components/icons.tsx         |   2 +
 .../settings/ModelSelectionPanes.tsx          |  31 ++++
 .../src/components/settings/provider-copy.ts  |   1 +
 .../chat/composer/ComposerModelPicker.tsx     |  15 +-
 .../chat/composer/ComposerToolbar.tsx         |   4 +
 apps/desktop/src/lib/api.ts                   |   2 +-
 apps/desktop/src/lib/commands.ts              |   4 +
 apps/desktop/src/lib/fast-mode-command.ts     |  78 +++++++++
 apps/desktop/src/lib/session-thinking.ts      |   1 +
 apps/desktop/src/stores/app-state.ts          |  12 +-
 .../stores/runtime/session-coordination.ts    |   1 +
 .../src/stores/runtime/session-runtime.ts     |   2 +-
 .../src/stores/slices/session-slice.ts        |   1 +
 apps/desktop/src/styles/composer-menus.css    |   8 +
 apps/desktop/test/builtin-commands.test.mjs   |   5 +-
 apps/desktop/test/fast-mode-command.test.mjs  | 155 ++++++++++++++++++
 .../test/fast-mode-session-launch.test.mjs    |  89 ++++++++++
 .../test/model-advanced-capabilities.test.mjs |  12 ++
 crates/host-core/src/db.rs                    |   4 +-
 crates/host-core/src/db/repositories.rs       |   7 +
 crates/host-core/src/db/schema.rs             |   1 +
 .../src/db/session_fast_mode_migration.rs     |  33 ++++
 crates/host-core/src/db/tests.rs              |  37 +++++
 crates/host-core/src/plugins/providers.rs     |   1 +
 .../host-core/src/plugins/providers/tests.rs  |   8 +-
 crates/host-core/src/providers/catalog.rs     |   3 +
 crates/host-core/src/providers/model.rs       |   4 +
 crates/host-core/src/providers/repository.rs  |   1 +
 crates/host-core/src/providers/tests.rs       |   8 +
 crates/host-core/src/rpc/mod.rs               |  72 +++++++-
 crates/host-core/src/rpc/scheduled_rpc.rs     |   2 +
 .../src/session_collaboration/mod.rs          |   2 +
 crates/host-core/src/session_search.rs        |   3 +-
 .../src/session_worktrees/delegate/tests.rs   |   1 +
 .../src/session_worktrees/lifecycle/tests.rs  |   2 +
 crates/host-core/src/sessions.rs              | 128 ++++++++++++++-
 crates/host-core/src/turn_queue.rs            |   2 +-
 docs/adr/0332-openai-fast-mode.md             |  69 ++++++++
 docs/adr/README.md                            |   1 +
 docs/project/unreleased.md                    |   6 +
 docs/spec/03-runtime/01-ipc-protocol.md       |   2 +
 docs/spec/03-runtime/04-data-storage.md       |  15 +-
 docs/spec/03-runtime/06-host-rpc-protocol.md  |   6 +-
 .../03-runtime/11-provider-model-system.md    |  12 ++
 docs/spec/04-ux/04-builtin-commands.md        |  23 ++-
 docs/spec/06-delivery/04-e2e-test-plan.md     |  31 ++++
 docs/zh-CN/spec/03-runtime/01-ipc-protocol.md |   2 +
 docs/zh-CN/spec/03-runtime/04-data-storage.md |  13 +-
 .../spec/03-runtime/06-host-rpc-protocol.md   |   6 +-
 .../03-runtime/11-provider-model-system.md    |   9 +
 docs/zh-CN/spec/04-ux/04-builtin-commands.md  |  18 +-
 .../spec/06-delivery/04-e2e-test-plan.md      |  18 ++
 packages/agent-runtime/src/fast-mode.test.ts  |  75 +++++++++
 packages/agent-runtime/src/fast-mode.ts       |  34 ++++
 packages/agent-runtime/src/runtime.ts         |  35 +++-
 packages/agent-runtime/src/sidecar.ts         |   4 +
 packages/i18n/src/locales/de/index.ts         |  11 ++
 packages/i18n/src/locales/en/index.ts         |  11 ++
 packages/i18n/src/locales/es/index.ts         |  11 ++
 packages/i18n/src/locales/fr/index.ts         |  11 ++
 packages/i18n/src/locales/ko/index.ts         |  11 ++
 packages/i18n/src/locales/pt-BR/index.ts      |  11 ++
 packages/i18n/src/locales/tr/index.ts         |  11 ++
 packages/i18n/src/locales/zh-CN/index.ts      |  11 ++
 packages/i18n/src/locales/zh-TW/index.ts      |  11 ++
 packages/shared/src/fast-mode.test.ts         |  19 +++
 packages/shared/src/fast-mode.ts              |  26 +++
 packages/shared/src/index.ts                  |   1 +
 packages/shared/src/model-config-import.ts    |   3 +
 packages/shared/src/provider-sync.ts          |   4 +-
 packages/shared/src/types/models.ts           |   6 +
 packages/shared/src/types/sessions.ts         |   6 +
 77 files changed, 1244 insertions(+), 57 deletions(-)
 create mode 100644 apps/desktop/src/lib/fast-mode-command.ts
 create mode 100644 apps/desktop/test/fast-mode-command.test.mjs
 create mode 100644 apps/desktop/test/fast-mode-session-launch.test.mjs
 create mode 100644 crates/host-core/src/db/session_fast_mode_migration.rs
 create mode 100644 docs/adr/0332-openai-fast-mode.md
 create mode 100644 packages/agent-runtime/src/fast-mode.test.ts
 create mode 100644 packages/agent-runtime/src/fast-mode.ts
 create mode 100644 packages/shared/src/fast-mode.test.ts
 create mode 100644 packages/shared/src/fast-mode.ts

diff --git a/apps/desktop/electron/main/builtin-commands.ts b/apps/desktop/electron/main/builtin-commands.ts
index 28ca880c80..e84cf36a3a 100644
--- a/apps/desktop/electron/main/builtin-commands.ts
+++ b/apps/desktop/electron/main/builtin-commands.ts
@@ -15,6 +15,7 @@ export const BUILTIN_COMMANDS: BuiltinCommandDef[] = [
   { id: "builtin.mode.agent", title: "Switch to Agent mode", category: "Session", keywords: ["mode", "agent"], source: "builtin", slash: "agent-mode" },
   { id: "builtin.mode.plan", title: "Switch to Plan mode", category: "Session", keywords: ["mode", "plan", "planning"], source: "builtin", slash: "plan-mode" },
   { id: "builtin.mode.goal", title: "Switch to Goal mode", category: "Session", keywords: ["mode", "goal", "objective", "autonomous"], source: "builtin", slash: "goal-mode" },
+  { id: "builtin.session.fast", title: "Toggle Fast mode", category: "Session", keywords: ["fast", "speed", "priority", "service tier"], source: "builtin", slash: "fast" },
 ];
 
 /** Palette-shaped items (no slash field leaks into the palette contract). */
diff --git a/apps/desktop/electron/main/ipc/session-ipc.ts b/apps/desktop/electron/main/ipc/session-ipc.ts
index f62be5b1e4..1fb56753a7 100644
--- a/apps/desktop/electron/main/ipc/session-ipc.ts
+++ b/apps/desktop/electron/main/ipc/session-ipc.ts
@@ -603,6 +603,7 @@ export function registerSessionIpc({
         modelId?: string;
         thinkingLevel?: SessionThinkingLevel;
         permissionMode?: "inherit" | "ask" | "accept-edits" | "auto";
+        fastMode?: boolean;
       },
     ) => {
       rejectNativeMutation(id, "configuration");
diff --git a/apps/desktop/electron/main/remote/remote-backend.ts b/apps/desktop/electron/main/remote/remote-backend.ts
index 0fc5d91c75..dd17b5029c 100644
--- a/apps/desktop/electron/main/remote/remote-backend.ts
+++ b/apps/desktop/electron/main/remote/remote-backend.ts
@@ -269,9 +269,14 @@ export function createRemoteBackend(options: RemoteBackendOptions): RemoteBacken
       case IPC.invoke.sessionConfigure: {
         const remoteSessionId = remoteIdFor(args);
         const config = (args[1] ?? {}) as Partial<
-          Pick<SessionSummary, "mode" | "providerId" | "modelId" | "thinkingLevel" | "permissionMode">
+          Pick<SessionSummary, "mode" | "providerId" | "modelId" | "thinkingLevel" | "permissionMode" | "fastMode">
         >;
-        if (config.providerId !== undefined || config.modelId !== undefined || config.thinkingLevel !== undefined) {
+        if (
+          config.providerId !== undefined ||
+          config.modelId !== undefined ||
+          config.thinkingLevel !== undefined ||
+          config.fastMode !== undefined
+        ) {
           throw capabilityUnavailable("remote sessions use the Host model configuration");
         }
         const { session } = await client.request<{ session: RacpSession }>("session/configure", {
diff --git a/apps/desktop/electron/main/runtime/session-launch.ts b/apps/desktop/electron/main/runtime/session-launch.ts
index 3461c2e709..0b82b13da2 100644
--- a/apps/desktop/electron/main/runtime/session-launch.ts
+++ b/apps/desktop/electron/main/runtime/session-launch.ts
@@ -706,6 +706,9 @@ export function createSessionLaunchRuntime({
         ...(overrides.turnId ? { turnId: overrides.turnId } : {}),
         thinkingLevel,
         infiniteProviderRetry: settings.infiniteProviderRetry === true,
+        // The session `/fast` switch only reaches a model the user allowed it
+        // for in model settings (ADR 0332).
+        fastMode: session.fastMode === true && storedModel?.fastModeAllowed === true,
         ...(jevApiKey ? { jevApiKey } : {}),
         commandShell,
         scratchDir: join(dataDir, "scratch", sessionId),
diff --git a/apps/desktop/src/components/Composer.tsx b/apps/desktop/src/components/Composer.tsx
index c9baf395d8..10f9b244de 100644
--- a/apps/desktop/src/components/Composer.tsx
+++ b/apps/desktop/src/components/Composer.tsx
@@ -25,6 +25,7 @@ import { isActivePlanExecution } from "../lib/plan-mode-state";
 import { headAsk, queuedAskCount } from "../lib/pending-asks";
 import type { QueuedPrompt } from "../lib/queued-prompts";
 import { composerModelDisplayName, sameComposerModelId } from "../lib/composer-models";
+import { sessionFastModeActive } from "../lib/fast-mode-command";
 import {
   resolveComposerThinkingProvider,
 } from "../lib/session-thinking";
@@ -626,6 +627,7 @@ export function Composer({
             modelMenu={modelMenu}
             modelLabel={modelLabel}
             thinkingLabel={thinkingLabel}
+            modelFastMode={sessionFastModeActive(provider, modelId, activeSession ?? draftConfiguration)}
             contextUsage={composerContextUsage ?? null}
             value={value}
             modelReady={modelReady}
diff --git a/apps/desktop/src/components/icons.tsx b/apps/desktop/src/components/icons.tsx
index 13f39d92c9..08337551f8 100644
--- a/apps/desktop/src/components/icons.tsx
+++ b/apps/desktop/src/components/icons.tsx
@@ -101,6 +101,7 @@ import {
   Workflow,
   Wrench,
   X,
+  Zap,
   type LucideIcon,
   type LucideProps,
 } from "lucide-react";
@@ -192,6 +193,7 @@ export const IconDatabase = icon(Database);
 export const IconCheck = icon(Check);
 export const IconBell = icon(Bell);
 export const IconBot = icon(Bot);
+export const IconZap = icon(Zap);
 export const IconCheckCheck = icon(CheckCheck);
 export const IconShield = icon(Shield);
 export const IconChevronDown = icon(ChevronDown);
diff --git a/apps/desktop/src/components/settings/ModelSelectionPanes.tsx b/apps/desktop/src/components/settings/ModelSelectionPanes.tsx
index 668a838b26..8b0192d9a4 100644
--- a/apps/desktop/src/components/settings/ModelSelectionPanes.tsx
+++ b/apps/desktop/src/components/settings/ModelSelectionPanes.tsx
@@ -16,6 +16,7 @@ import {
   bindingForCustomModel,
   bindingForCustomModelInfo,
   bindingFromModelInfo,
+  fastModeBillingOn,
   formatTokenCount,
   modelMatchesFilter,
   nativeWebSearchSupportedOn,
@@ -307,6 +308,14 @@ export function ModelSelectionPanes({
   // Use the same published endpoint routing as the runtime. A disabled control
   // means this connection is not integrated, not that the vendor cannot search.
   const nativeWebSearchWireCapable = nativeWebSearchSupportedOn(apiStyle, lookupContext?.baseUrl);
+  // Fast mode rides the Responses `service_tier` field; the hint names what
+  // the priority tier costs on this wire (ADR 0332).
+  const fastModeBilling = fastModeBillingOn(apiStyle);
+  const fastModeHint = fastModeBilling === "subscription"
+    ? "settings.fastModeAllowedSubscriptionHint"
+    : fastModeBilling === "usage"
+      ? "settings.fastModeAllowedUsageHint"
+      : "settings.fastModeAllowedUnsupported";
 
   /**
    * The chosen list narrows with the discovered list's rule plus the binding's
@@ -1032,6 +1041,28 @@ export function ModelSelectionPanes({
                             <IconHelp size={13} />
                           </Tooltip>
                         </span>
+                        <span className="provider-chosen-delegation">
+                          <label className="provider-chosen-capability">
+                            <input
+                              type="checkbox"
+                              checked={binding.fastModeAllowed === true}
+                              disabled={!fastModeBilling}
+                              onChange={(event) =>
+                                updateBinding(binding.id, {
+                                  fastModeAllowed: event.target.checked || undefined,
+                                })
+                              }
+                            />
+                            <span>{t("settings.fastModeAllowed")}</span>
+                          </label>
+                          <Tooltip
+                            className="provider-chosen-delegation-help"
+                            label={t(fastModeHint)}
+                            ariaLabel={t(fastModeHint)}
+                          >
+                            <IconHelp size={13} />
+                          </Tooltip>
+                        </span>
                       </div>
                     </div>
                   </div>
diff --git a/apps/desktop/src/components/settings/provider-copy.ts b/apps/desktop/src/components/settings/provider-copy.ts
index 54b06748bd..eb3805547a 100644
--- a/apps/desktop/src/components/settings/provider-copy.ts
+++ b/apps/desktop/src/components/settings/provider-copy.ts
@@ -47,6 +47,7 @@ export function copyProviderConfiguration(provider: ProviderPublic, name: string
       ...(model.supportsDocuments !== undefined ? { supportsDocuments: model.supportsDocuments } : {}),
       ...(model.availableForSubagents !== undefined ? { availableForSubagents: model.availableForSubagents } : {}),
       ...(model.nativeWebSearch !== undefined ? { nativeWebSearch: model.nativeWebSearch } : {}),
+      ...(model.fastModeAllowed !== undefined ? { fastModeAllowed: model.fastModeAllowed } : {}),
     })),
   };
 }
diff --git a/apps/desktop/src/features/chat/composer/ComposerModelPicker.tsx b/apps/desktop/src/features/chat/composer/ComposerModelPicker.tsx
index 05a89f8158..d9e9431981 100644
--- a/apps/desktop/src/features/chat/composer/ComposerModelPicker.tsx
+++ b/apps/desktop/src/features/chat/composer/ComposerModelPicker.tsx
@@ -2,7 +2,7 @@ import type { TFunction } from "i18next";
 import type { ReactNode } from "react";
 import { ComposerModelList } from "./ComposerModelList";
 import { AnchoredMenu } from "../../../components/settings/AnchoredMenu";
-import { IconBot, IconChevronDown } from "../../../components/icons";
+import { IconBot, IconChevronDown, IconZap } from "../../../components/icons";
 import { TooltipButton } from "../../../components/ui";
 import type { useComposerModelMenu } from "./hooks/useComposerModelMenu";
 import { ThinkingLevelSlider } from "./ThinkingLevelSlider";
@@ -15,6 +15,8 @@ export type ComposerModelPickerProps = {
   modelLabel: string;
   thinkingLabel: string;
   thinkingLevel: string;
+  /** The selected model asks for OpenAI's priority tier (ADR 0332). */
+  fastMode?: boolean;
   selectedProviderId?: string;
   selectedModelId?: string;
   controlsBlocked: boolean;
@@ -29,6 +31,7 @@ export function ComposerModelPicker({
   modelLabel,
   thinkingLabel,
   thinkingLevel,
+  fastMode = false,
   selectedProviderId,
   selectedModelId,
   controlsBlocked,
@@ -54,6 +57,7 @@ export function ComposerModelPicker({
     commitThinkingLevel,
     onMenuKeyDown,
   } = controller;
+  const fastSuffix = fastMode ? ` · ${t("chat.fastModeBadge")}` : "";
 
   return (
     <AnchoredMenu
@@ -72,8 +76,8 @@ export function ComposerModelPicker({
           ref={ref}
           type="button"
           className={`icon-btn composer-model-thinking-chip ${open ? "active" : ""}`}
-          tooltip={`${modelLabel} · ${t("chat.reasoningLevel")}: ${thinkingLabel}`}
-          ariaLabel={`${t("chat.model")}: ${modelLabel}. ${t("chat.reasoningLevel")}: ${thinkingLabel}`}
+          tooltip={`${modelLabel} · ${t("chat.reasoningLevel")}: ${thinkingLabel}${fastSuffix}`}
+          ariaLabel={`${t("chat.model")}: ${modelLabel}. ${t("chat.reasoningLevel")}: ${thinkingLabel}${fastSuffix}`}
           aria-haspopup="menu"
           aria-expanded={open}
           disabled={controlsBlocked}
@@ -94,6 +98,11 @@ export function ComposerModelPicker({
           {thinkingLevel !== "off" ? (
             <span className="composer-model-thinking-level">{thinkingLabel}</span>
           ) : null}
+          {fastMode ? (
+            <span className="composer-model-thinking-fast" aria-hidden="true">
+              <IconZap size={12} />
+            </span>
+          ) : null}
           <IconChevronDown size={12} aria-hidden="true" className="composer-model-thinking-chevron" />
         </TooltipButton>
       )}
diff --git a/apps/desktop/src/features/chat/composer/ComposerToolbar.tsx b/apps/desktop/src/features/chat/composer/ComposerToolbar.tsx
index 2b04f19f85..526ab6400e 100644
--- a/apps/desktop/src/features/chat/composer/ComposerToolbar.tsx
+++ b/apps/desktop/src/features/chat/composer/ComposerToolbar.tsx
@@ -51,6 +51,8 @@ export type ComposerToolbarProps = {
   modelMenu: ModelMenuController;
   modelLabel: string;
   thinkingLabel: string;
+  /** The selected model asks for OpenAI's priority tier (ADR 0332). */
+  modelFastMode?: boolean;
   contextUsage: ContextUsage | null;
   value: string;
   modelReady: boolean;
@@ -92,6 +94,7 @@ export function ComposerToolbar({
   modelMenu,
   modelLabel,
   thinkingLabel,
+  modelFastMode = false,
   contextUsage,
   value,
   modelReady,
@@ -189,6 +192,7 @@ export function ComposerToolbar({
           controller={modelMenu}
           modelLabel={modelLabel}
           thinkingLabel={thinkingLabel}
+          fastMode={modelFastMode}
           thinkingLevel={thinkingLevel}
           selectedProviderId={providerId}
           selectedModelId={modelId}
diff --git a/apps/desktop/src/lib/api.ts b/apps/desktop/src/lib/api.ts
index a23632a115..5d43fb6393 100644
--- a/apps/desktop/src/lib/api.ts
+++ b/apps/desktop/src/lib/api.ts
@@ -760,7 +760,7 @@ export const api = {
   configureSession: (
     id: string,
     config: Pick<SessionSummary, "mode" | "providerId" | "modelId"> &
-      Partial<Pick<SessionSummary, "thinkingLevel" | "permissionMode">>,
+      Partial<Pick<SessionSummary, "thinkingLevel" | "permissionMode" | "fastMode">>,
   ) =>
     invoke<{ session: SessionSummary }>(
       IPC.invoke.sessionConfigure,
diff --git a/apps/desktop/src/lib/commands.ts b/apps/desktop/src/lib/commands.ts
index f399cdb8df..081016e129 100644
--- a/apps/desktop/src/lib/commands.ts
+++ b/apps/desktop/src/lib/commands.ts
@@ -1,6 +1,7 @@
 import i18n from "i18next";
 import { api } from "./api";
 import { useAppStore } from "../stores/app-store";
+import { toggleSessionFastMode } from "./fast-mode-command";
 import { trustedExtensionCommandName, type Mode } from "@pi-desktop/shared";
 
 /**
@@ -17,6 +18,9 @@ export async function runPaletteCommand(commandId: string): Promise<void> {
     case "builtin.agent.compact":
       await store.compactContext();
       break;
+    case "builtin.session.fast":
+      await toggleSessionFastMode();
+      break;
     case "builtin.mode.agent":
     case "builtin.mode.plan":
     case "builtin.mode.goal": {
diff --git a/apps/desktop/src/lib/fast-mode-command.ts b/apps/desktop/src/lib/fast-mode-command.ts
new file mode 100644
index 0000000000..6fefb5763e
--- /dev/null
+++ b/apps/desktop/src/lib/fast-mode-command.ts
@@ -0,0 +1,78 @@
+import i18n from "i18next";
+import { fastModeBillingOn, type ProviderPublic } from "@pi-desktop/shared";
+import { useAppStore } from "../stores/app-store";
+import { sameComposerModelId } from "./composer-models";
+import { isRemoteSession } from "./remote-session-safety";
+import { inheritedSessionModelBinding } from "./session-model";
+
+/**
+ * The `/fast` builtin (ADR 0332): a per-session switch for OpenAI's priority
+ * service tier. Model settings decide which models may use it.
+ */
+
+type FastModeProvider = Pick<ProviderPublic, "apiStyle" | "models"> | undefined | null;
+
+function fastModeAllowed(provider: FastModeProvider, modelId: string | undefined): boolean {
+  return !!modelId && provider?.models.some(
+    (binding) => sameComposerModelId(binding.id, modelId) && binding.fastModeAllowed === true,
+  ) === true;
+}
+
+/** Whether the Composer should mark the selected model as running in Fast mode. */
+export function sessionFastModeActive(
+  provider: FastModeProvider,
+  modelId: string | undefined,
+  target: { fastMode?: boolean } | null | undefined,
+): boolean {
+  return target?.fastMode === true && !!fastModeBillingOn(provider?.apiStyle) &&
+    fastModeAllowed(provider, modelId);
+}
+
+export async function toggleSessionFastMode(): Promise<void> {
+  const store = useAppStore.getState();
+  const activeSession = store.sessions.find((session) => session.id === store.activeSessionId);
+  if (store.activeSessionId && !activeSession) return;
+  if (activeSession && isRemoteSession(activeSession)) {
+    store.showToast(i18n.t("chat.fastModeRemote"), { variant: "warning" });
+    return;
+  }
+  const target = activeSession ?? store.draftConfiguration;
+  const enabled = target?.fastMode !== true;
+  const selected = inheritedSessionModelBinding({
+    draft: target,
+    settings: store.settings,
+    providers: store.providers,
+    recentModels: store.recentModels,
+  });
+  const provider = store.providers.find((candidate) => candidate.id === selected.providerId);
+  const billing = fastModeBillingOn(provider?.apiStyle);
+  // Without a configured model the switch does nothing, so `/fast` only
+  // points to the setting; a stale stored switch stays inert at launch.
+  if (!billing) {
+    store.showToast(i18n.t("chat.fastModeUnsupported"), { variant: "warning" });
+    return;
+  }
+  if (!fastModeAllowed(provider, selected.modelId)) {
+    store.showToast(i18n.t("chat.fastModeNotAllowed"), { variant: "warning" });
+    return;
+  }
+
+  if (activeSession) {
+    await store.configureActiveSession({
+      mode: activeSession.mode,
+      providerId: activeSession.providerId,
+      modelId: activeSession.modelId,
+      thinkingLevel: activeSession.thinkingLevel,
+      fastMode: enabled,
+    });
+    // A pending plan review or a read-only session leaves the switch as is.
+    const after = useAppStore.getState().sessions.find((session) => session.id === activeSession.id);
+    if ((after?.fastMode === true) !== enabled) return;
+  } else {
+    useAppStore.setState((state) => ({
+      draftConfiguration: { ...state.draftConfiguration, fastMode: enabled },
+    }));
+  }
+  const message = !enabled ? "chat.fastModeOff" : billing === "usage" ? "chat.fastModeOnUsage" : "chat.fastModeOn";
+  store.showToast(i18n.t(message), { variant: "info" });
+}
diff --git a/apps/desktop/src/lib/session-thinking.ts b/apps/desktop/src/lib/session-thinking.ts
index d4b4655e76..a9294f6d5f 100644
--- a/apps/desktop/src/lib/session-thinking.ts
+++ b/apps/desktop/src/lib/session-thinking.ts
@@ -26,6 +26,7 @@ export type OptimisticSessionConfiguration = {
   thinkingLevel?: SessionThinkingLevel;
   mode?: unknown;
   permissionMode?: unknown;
+  fastMode?: boolean;
 };
 
 export type ThinkingProviderLike = {
diff --git a/apps/desktop/src/stores/app-state.ts b/apps/desktop/src/stores/app-state.ts
index 21b76b5b9c..3319a148b6 100644
--- a/apps/desktop/src/stores/app-state.ts
+++ b/apps/desktop/src/stores/app-state.ts
@@ -104,13 +104,18 @@ export type RefreshSessionsOptions = {
   revealImportedProjects?: boolean;
 };
 
-/** Toolbar selections retained on the unpersisted new-task draft. */
+/**
+ * Toolbar selections retained on the unpersisted new-task draft. An absent
+ * field keeps the default the new session would otherwise start with.
+ */
 export type DraftSessionConfiguration = {
-  mode: Mode;
-  thinkingLevel: SessionThinkingLevel;
+  mode?: Mode;
+  thinkingLevel?: SessionThinkingLevel;
   providerId?: string;
   modelId?: string;
   permissionMode?: PermissionMode;
+  /** Session `/fast` switch carried into the created session (ADR 0332). */
+  fastMode?: boolean;
 };
 
 
@@ -228,6 +233,7 @@ export type AppState = {
     modelId?: string;
     thinkingLevel: SessionThinkingLevel;
     permissionMode?: PermissionMode;
+    fastMode?: boolean;
   }) => Promise<void>;
   /** Returns true once accepted unless concurrent smart Stop restores it. */
   sendPrompt: (
diff --git a/apps/desktop/src/stores/runtime/session-coordination.ts b/apps/desktop/src/stores/runtime/session-coordination.ts
index 3e9fefe368..c79d263fdd 100644
--- a/apps/desktop/src/stores/runtime/session-coordination.ts
+++ b/apps/desktop/src/stores/runtime/session-coordination.ts
@@ -319,6 +319,7 @@ export function createSessionCoordination({
         mode: draftConfig?.mode ?? normalizeMode(settings?.defaultMode),
         thinkingLevel: draftConfig?.thinkingLevel ?? defaultThinkingLevel,
         permissionMode: draftConfig?.permissionMode,
+        ...(draftConfig?.fastMode ? { fastMode: true } : {}),
         providerId: inherited.providerId,
         modelId: inherited.modelId,
         projectPath: projectPath ?? undefined,
diff --git a/apps/desktop/src/stores/runtime/session-runtime.ts b/apps/desktop/src/stores/runtime/session-runtime.ts
index 26dbb620e4..b879617d99 100644
--- a/apps/desktop/src/stores/runtime/session-runtime.ts
+++ b/apps/desktop/src/stores/runtime/session-runtime.ts
@@ -43,7 +43,7 @@ export type SessionConfiguration = Pick<
   SessionSummary,
   "mode" | "providerId" | "modelId" | "thinkingLevel"
 > &
-  Partial<Pick<SessionSummary, "permissionMode">>;
+  Partial<Pick<SessionSummary, "permissionMode" | "fastMode">>;
 
 export type SessionSelection = { id: string; intent: number };
 
diff --git a/apps/desktop/src/stores/slices/session-slice.ts b/apps/desktop/src/stores/slices/session-slice.ts
index e1c41d699b..5da762f07b 100644
--- a/apps/desktop/src/stores/slices/session-slice.ts
+++ b/apps/desktop/src/stores/slices/session-slice.ts
@@ -790,6 +790,7 @@ export function createSessionSlice({
             modelId: config.modelId ?? state.draftConfiguration?.modelId,
             permissionMode:
               config.permissionMode ?? state.draftConfiguration?.permissionMode,
+            fastMode: config.fastMode ?? state.draftConfiguration?.fastMode,
           },
         }));
         return;
diff --git a/apps/desktop/src/styles/composer-menus.css b/apps/desktop/src/styles/composer-menus.css
index 5fdfe01df9..2c67aec382 100644
--- a/apps/desktop/src/styles/composer-menus.css
+++ b/apps/desktop/src/styles/composer-menus.css
@@ -189,6 +189,14 @@
   line-height: var(--leading-compact);
 }
 
+/* Fast mode stays visible at every width: it changes what each turn costs. */
+.composer-model-thinking-fast {
+  flex: 0 0 auto;
+  display: inline-flex;
+  align-items: center;
+  color: var(--ds-text-secondary);
+}
+
 .composer-model-thinking-chevron {
   flex: 0 0 auto;
   color: var(--ds-text-faint);
diff --git a/apps/desktop/test/builtin-commands.test.mjs b/apps/desktop/test/builtin-commands.test.mjs
index 477cd6c7f2..9e69d7734f 100644
--- a/apps/desktop/test/builtin-commands.test.mjs
+++ b/apps/desktop/test/builtin-commands.test.mjs
@@ -17,9 +17,10 @@ const coreIds = [
   "builtin.mode.agent",
   "builtin.mode.plan",
   "builtin.mode.goal",
+  "builtin.session.fast",
 ];
 
-const coreAliases = ["new", "compact", "agent-mode", "plan-mode", "goal-mode"];
+const coreAliases = ["new", "compact", "agent-mode", "plan-mode", "goal-mode", "fast"];
 
 const removedIds = [
   "builtin.session.delete",
@@ -38,7 +39,7 @@ const removedIds = [
   "builtin.app.toggleDevtools",
 ];
 
-test("builtin registry exposes exactly the five core commands and aliases", () => {
+test("builtin registry exposes exactly the six core commands and aliases", () => {
   const ids = [...registry.matchAll(/\bid: "([^"]+)"/g)].map((match) => match[1]);
   const aliases = [...registry.matchAll(/slash: "([^"]+)"/g)].map((match) => match[1]);
 
diff --git a/apps/desktop/test/fast-mode-command.test.mjs b/apps/desktop/test/fast-mode-command.test.mjs
new file mode 100644
index 0000000000..63bd460f45
--- /dev/null
+++ b/apps/desktop/test/fast-mode-command.test.mjs
@@ -0,0 +1,155 @@
+import i18n from "i18next";
+import { en } from "@pi-desktop/i18n";
+import assert from "node:assert/strict";
+import { register } from "node:module";
+import test from "node:test";
+register(new URL("./helpers/ts-import-hooks.mjs", import.meta.url));
+const { useAppStore, materializeDraftSession } = await import("../src/stores/app-store.ts");
+const { api } = await import("../src/lib/api.ts");
+const { runPaletteCommand } = await import("../src/lib/commands.ts");
+const { sessionFastModeActive } = await import("../src/lib/fast-mode-command.ts");
+
+// ADR 0332: `/fast` is a per-session switch, and model settings decide which
+// models may use it.
+
+await i18n.init({ lng: "en", keySeparator: false, resources: { en: { translation: flatten(en) } } });
+
+function flatten(catalog, prefix = "", out = {}) {
+  for (const [key, value] of Object.entries(catalog)) {
+    if (value && typeof value === "object") flatten(value, `${prefix}${key}.`, out);
+    else out[`${prefix}${key}`] = value;
+  }
+  return out;
+}
+
+const codex = {
+  id: "codex", name: "ChatGPT", enabled: true, hasSecret: true, authKind: "oauth", apiStyle: "openai_codex_responses",
+  models: [{ id: "gpt-codex", fastModeAllowed: true }, { id: "gpt-mini" }],
+};
+const relay = {
+  id: "relay", name: "Relay", enabled: true, hasSecret: true, authKind: "api_key", apiStyle: "responses",
+  models: [{ id: "gpt-relay", fastModeAllowed: true }],
+};
+const completions = {
+  id: "completions", name: "Gateway", enabled: true, hasSecret: true, authKind: "api_key", apiStyle: "chat_completions",
+  models: [{ id: "chat", fastModeAllowed: true }],
+};
+const session = (provider, modelId, extra = {}) => ({
+  id: "s1", title: "Task", messageCount: 0, mode: "agent", thinkingLevel: "medium", permissionMode: "inherit",
+  providerId: provider.id, modelId, createdAt: "2026-01-01T00:00:00Z", updatedAt: "2026-01-01T00:00:00Z", ...extra,
+});
+
+function setup(t, state) {
+  const toasts = [];
+  const configures = [];
+  const creates = [];
+  t.mock.method(api, "configureSession", async (id, config) => {
+    configures.push(config);
+    return { session: { ...useAppStore.getState().sessions.find((item) => item.id === id), ...config, id } };
+  });
+  t.mock.method(api, "createSession", async (config) => {
+    creates.push(config);
+    return { session: { ...config, id: "created", title: "New task", messageCount: 0, messages: [] } };
+  });
+  useAppStore.setState({
+    providers: [codex, relay, completions],
+    sessions: [],
+    activeSessionId: null,
+    runningSessions: {},
+    pendingPlans: {},
+    recentModels: [],
+    draftConfiguration: null,
+    workspace: null,
+    settings: { defaultProviderId: codex.id, defaultModelId: "gpt-codex", defaultMode: "agent" },
+    showToast: (message, options) => { toasts.push({ message, variant: options?.variant }); },
+    ...state,
+  });
+  return { toasts, configures, creates };
+}
+
+test("/fast toggles an allowed subscription model's session through session.configure", async (t) => {
+  const { toasts, configures } = setup(t, { sessions: [session(codex, "gpt-codex")], activeSessionId: "s1" });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures.length, 1);
+  assert.equal(configures[0].fastMode, true);
+  assert.equal(configures[0].thinkingLevel, "medium", "other selections are resent unchanged");
+  assert.equal(useAppStore.getState().sessions[0].fastMode, true);
+  assert.equal(toasts[0].message, en.chat.fastModeOn);
+
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures[1].fastMode, false);
+  assert.equal(useAppStore.getState().sessions[0].fastMode, false);
+  assert.equal(toasts[1].message, en.chat.fastModeOff);
+});
+
+test("/fast on a per-token Responses model warns about the price", async (t) => {
+  const { toasts, configures } = setup(t, { sessions: [session(relay, "gpt-relay")], activeSessionId: "s1" });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures[0].fastMode, true);
+  assert.equal(toasts[0].message, en.chat.fastModeOnUsage);
+});
+
+test("/fast refuses a model the user has not allowed, and a wire without the tier", async (t) => {
+  const notAllowed = setup(t, { sessions: [session(codex, "gpt-mini")], activeSessionId: "s1" });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(notAllowed.configures.length, 0);
+  assert.deepEqual(notAllowed.toasts, [{ message: en.chat.fastModeNotAllowed, variant: "warning" }]);
+
+  const unsupported = setup(t, { sessions: [session(completions, "chat")], activeSessionId: "s1" });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(unsupported.configures.length, 0);
+  assert.deepEqual(unsupported.toasts, [{ message: en.chat.fastModeUnsupported, variant: "warning" }]);
+});
+
+test("/fast points to model settings instead of toggling a stale switch on an unconfigured model", async (t) => {
+  const { configures, toasts } = setup(t, {
+    sessions: [session(codex, "gpt-mini", { fastMode: true })],
+    activeSessionId: "s1",
+  });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures.length, 0);
+  assert.deepEqual(toasts, [{ message: en.chat.fastModeNotAllowed, variant: "warning" }]);
+});
+
+test("/fast refuses remote sessions", async (t) => {
+  const { configures, toasts } = setup(t, {
+    sessions: [session(codex, "gpt-codex", { id: "remote:host:s1", source: "remote" })],
+    activeSessionId: "remote:host:s1",
+  });
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures.length, 0);
+  assert.equal(toasts[0].message, en.chat.fastModeRemote);
+});
+
+test("/fast on the new-task draft carries the switch into the created session", async (t) => {
+  const { configures, creates } = setup(t, {});
+  await runPaletteCommand("builtin.session.fast");
+  assert.equal(configures.length, 0);
+  assert.deepEqual(useAppStore.getState().draftConfiguration, { fastMode: true });
+
+  // Choosing a model on the draft keeps the switch.
+  await useAppStore.getState().configureActiveSession({
+    mode: "agent", providerId: codex.id, modelId: "gpt-codex", thinkingLevel: "high",
+  });
+  assert.equal(useAppStore.getState().draftConfiguration.fastMode, true);
+
+  await materializeDraftSession();
+  assert.equal(creates.length, 1);
+  assert.equal(creates[0].fastMode, true);
+  assert.equal(creates[0].thinkingLevel, "high");
+});
+
+test("a draft without /fast creates a session on the default tier and default thinking", async (t) => {
+  const { creates } = setup(t, {});
+  await materializeDraftSession();
+  assert.equal(creates.length, 1);
+  assert.equal("fastMode" in creates[0], false);
+});
+
+test("the Composer marks Fast only when the switch is on for an allowed model", () => {
+  assert.equal(sessionFastModeActive(codex, "gpt-codex", { fastMode: true }), true);
+  assert.equal(sessionFastModeActive(codex, "gpt-codex", { fastMode: false }), false);
+  assert.equal(sessionFastModeActive(codex, "gpt-mini", { fastMode: true }), false);
+  assert.equal(sessionFastModeActive(completions, "chat", { fastMode: true }), false);
+  assert.equal(sessionFastModeActive(undefined, undefined, { fastMode: true }), false);
+});
diff --git a/apps/desktop/test/fast-mode-session-launch.test.mjs b/apps/desktop/test/fast-mode-session-launch.test.mjs
new file mode 100644
index 0000000000..139955ffbb
--- /dev/null
+++ b/apps/desktop/test/fast-mode-session-launch.test.mjs
@@ -0,0 +1,89 @@
+import assert from "node:assert/strict";
+import { createRequire } from "node:module";
+import { mkdtemp, rm } from "node:fs/promises";
+import { join } from "node:path";
+import test from "node:test";
+import { pathToFileURL } from "node:url";
+import { repositoryRoot } from "../../../scripts/e2e/boot.mjs";
+
+const root = repositoryRoot();
+const { build } = createRequire(join(root, "packages/agent-runtime/package.json"))("esbuild");
+
+test("session launch sends Fast mode only for a model the user allowed it on (ADR 0332)", async () => {
+  const temp = await mkdtemp(join(root, "apps/desktop/.fast-mode-launch-test-"));
+  const priorHome = process.env.HOME;
+  process.env.HOME = temp;
+  try {
+    const bundle = join(temp, "session-launch.mjs");
+    await build({
+      entryPoints: [join(root, "apps/desktop/electron/main/runtime/session-launch.ts")],
+      outfile: bundle,
+      bundle: true,
+      platform: "node",
+      format: "esm",
+      packages: "external",
+      nodePaths: [join(root, "apps/desktop/node_modules")],
+    });
+    const { createSessionLaunchRuntime } = await import(pathToFileURL(bundle).href);
+    const provider = {
+      id: "openai-provider",
+      name: "OpenAI",
+      authKind: "none",
+      apiStyle: "responses",
+      baseUrl: "http://127.0.0.1:11434/v1",
+      enabled: true,
+      defaultModelId: "allowed-model",
+      models: [],
+    };
+    const allowed = new Set(["allowed-model"]);
+    const host = {
+      isAvailable: () => true,
+      async call(method) {
+        if (method === "commandShells.list") {
+          const shell = { id: "bash", label: "Bash", dialect: "posix", available: true, isDefault: true };
+          return { configuredId: "bash", effective: shell, fallback: false, choices: [shell] };
+        }
+        if (method === "providers.list") return { providers: [provider] };
+        if (method === "providers.getSecret") return { value: "" };
+        if (method === "skills.disabledBuiltins") return { disabled: [] };
+        if (method === "skills.active") return { skills: [] };
+        if (method === "mcp.active") return { servers: [] };
+        if (method === "agents.active") return { subagents: [] };
+        if (method === "agents.disabledBuiltins") return { disabled: [] };
+        throw new Error("Unexpected fixture Host method: " + method);
+      },
+    };
+    const launchRuntime = createSessionLaunchRuntime({
+      runtimeState: { host, sidecar: { setVendorAuthBindings() {} }, agentHostBridge: null },
+      logger: { app() {} },
+      userMcp: { setRecords() {}, async toolsForProject() { return []; } },
+      plugins: { getTools: () => [], getSkills: () => [], listLoaded: () => [], getAgentExtensions: () => [] },
+      sessionProjects: new Map(),
+      dataDir: temp,
+      vendorOAuth: { async bindingFor() { return undefined; } },
+      modelsDevCatalog: { async ensureLoaded() {}, configureAccount() {}, findModel() { return undefined; } },
+      getWorkspacePath: () => null,
+      pluginActiveInProject: () => true,
+      bindingForModel: (_provider, modelId) =>
+        allowed.has(modelId) ? { id: modelId, fastModeAllowed: true } : { id: modelId },
+      effectiveSubagentModelConfig: () => ({ modelConfig: undefined, capabilities: {} }),
+      normalizeThinkingLevel: () => "off",
+    });
+    const launch = async (modelId, fastMode) => {
+      const { sidecarParams } = await launchRuntime.resolveAgentRuntimeLaunch(
+        "session-fixture",
+        { providerId: provider.id, modelId, ...(fastMode === undefined ? {} : { fastMode }) },
+        { defaultProviderId: provider.id, defaultModelId: modelId, defaultMode: "agent" },
+      );
+      return sidecarParams.fastMode;
+    };
+
+    assert.equal(await launch("allowed-model", true), true);
+    assert.equal(await launch("allowed-model", false), false);
+    assert.equal(await launch("allowed-model", undefined), false);
+    assert.equal(await launch("other-model", true), false);
+  } finally {
+    process.env.HOME = priorHome;
+    await rm(temp, { recursive: true, force: true });
+  }
+});
diff --git a/apps/desktop/test/model-advanced-capabilities.test.mjs b/apps/desktop/test/model-advanced-capabilities.test.mjs
index 8e1010564a..fb2ab62848 100644
--- a/apps/desktop/test/model-advanced-capabilities.test.mjs
+++ b/apps/desktop/test/model-advanced-capabilities.test.mjs
@@ -285,3 +285,15 @@ test("the offered default and the saved default use one order", () => {
     /resolveBindingDefaultThinkingLevel\(\s*binding\.defaultThinkingLevel,\s*sortThinkingLevels\(next\),/,
   );
 });
+
+test("model settings let the user allow Fast mode per model on Responses wires", () => {
+  // ADR 0332: the opt-in lives on the binding; `/fast` is the per-session switch.
+  assert.match(pickerSource, /const fastModeBilling = fastModeBillingOn\(apiStyle\);/);
+  assert.match(pickerSource, /checked=\{binding\.fastModeAllowed === true\}/);
+  assert.match(pickerSource, /disabled=\{!fastModeBilling\}/);
+  assert.match(pickerSource, /fastModeAllowed: event\.target\.checked \|\| undefined/);
+  // Plan limits and per-token billing are different costs, so the hint differs.
+  assert.match(pickerSource, /"settings\.fastModeAllowedSubscriptionHint"/);
+  assert.match(pickerSource, /"settings\.fastModeAllowedUsageHint"/);
+  assert.match(pickerSource, /"settings\.fastModeAllowedUnsupported"/);
+});
diff --git a/crates/host-core/src/db.rs b/crates/host-core/src/db.rs
index 518e0005a5..7511d4ae77 100644
--- a/crates/host-core/src/db.rs
+++ b/crates/host-core/src/db.rs
@@ -7,7 +7,7 @@ pub(crate) use serde_json::Value;
 pub(crate) use std::path::{Path, PathBuf};
 
 /// Current SQLite schema version.
-pub const SCHEMA_VERSION: i64 = 24;
+pub const SCHEMA_VERSION: i64 = 25;
 
 /// Absolute approval deadline for a newly submitted Plan or Goal proposal.
 pub const PLAN_APPROVAL_TIMEOUT_MS: i64 = 30 * 60 * 1000;
@@ -22,6 +22,7 @@ mod project_groups;
 mod repositories;
 mod schema;
 mod session_collaboration_migration;
+mod session_fast_mode_migration;
 mod session_title_source_migration;
 mod worktree_migration;
 pub(crate) use migrations::{
@@ -43,6 +44,7 @@ pub(crate) use repositories::{
     OBSOLETE_PLAN_APPROVAL_PERMISSION_MODE,
 };
 pub(crate) use schema::{PLAN_APPROVALS_SCHEMA, SCHEMA_LATEST, SESSION_TODO_DDL};
+pub(crate) use session_fast_mode_migration::migrate_v24_to_v25;
 pub(crate) use session_title_source_migration::migrate_v22_to_v23;
 pub(crate) use worktree_migration::{migrate_v23_to_v24, WORKTREE_DDL};
 
diff --git a/crates/host-core/src/db/repositories.rs b/crates/host-core/src/db/repositories.rs
index 30bc39d30d..2f5ee3a0ff 100644
--- a/crates/host-core/src/db/repositories.rs
+++ b/crates/host-core/src/db/repositories.rs
@@ -207,6 +207,9 @@ impl Database {
             23 => {
                 migrate_v23_to_v24(&conn, path)?;
             }
+            24 => {
+                migrate_v24_to_v25(&conn, path)?;
+            }
             legacy @ 1..=6 => {
                 let _ = conn.execute_batch("PRAGMA wal_checkpoint(TRUNCATE);");
                 drop(conn);
@@ -257,6 +260,10 @@ impl Database {
         }
         if migrated_version == 23 {
             migrate_v23_to_v24(&conn, path)?;
+            migrated_version = conn.query_row("PRAGMA user_version", [], |row| row.get(0))?;
+        }
+        if migrated_version == 24 {
+            migrate_v24_to_v25(&conn, path)?;
         }
         let db = Self { conn, data_dir };
         db.boot_maintenance()?;
diff --git a/crates/host-core/src/db/schema.rs b/crates/host-core/src/db/schema.rs
index 2237f05190..8023a534f6 100644
--- a/crates/host-core/src/db/schema.rs
+++ b/crates/host-core/src/db/schema.rs
@@ -98,6 +98,7 @@ CREATE TABLE sessions (
                                           'high', 'xhigh', 'max', 'omit')),
   permission_mode TEXT NOT NULL DEFAULT 'inherit'
                 CHECK (permission_mode IN ('inherit', 'ask', 'accept-edits', 'auto')),
+  fast_mode   INTEGER NOT NULL DEFAULT 0 CHECK (fast_mode IN (0, 1)),
   source      TEXT,
   deleted_at  INTEGER,
   pinned      INTEGER NOT NULL DEFAULT 0,
diff --git a/crates/host-core/src/db/session_fast_mode_migration.rs b/crates/host-core/src/db/session_fast_mode_migration.rs
new file mode 100644
index 0000000000..e24503e6f3
--- /dev/null
+++ b/crates/host-core/src/db/session_fast_mode_migration.rs
@@ -0,0 +1,33 @@
+use super::{create_migration_backup, Connection, Context, Path, Result};
+use rusqlite::Transaction;
+
+/// v25 stores the per-session `/fast` switch (ADR 0332). Every existing
+/// session keeps the default service tier.
+fn migrate_v24_to_v25_tx(tx: &Transaction<'_>) -> Result<()> {
+    let has_fast_mode: bool = tx.query_row(
+        "SELECT EXISTS(SELECT 1 FROM pragma_table_info('sessions') WHERE name = 'fast_mode')",
+        [],
+        |row| row.get(0),
+    )?;
+    if !has_fast_mode {
+        tx.execute_batch(
+            "ALTER TABLE sessions ADD COLUMN fast_mode INTEGER NOT NULL DEFAULT 0
+             CHECK (fast_mode IN (0, 1));",
+        )?;
+    }
+    tx.pragma_update(None, "user_version", 25i64)?;
+    Ok(())
+}
+
+pub(crate) fn migrate_v24_to_v25(conn: &Connection, path: &Path) -> Result<()> {
+    let backup = create_migration_backup(conn, path, 24)?;
+    let tx = conn.unchecked_transaction()?;
+    migrate_v24_to_v25_tx(&tx)?;
+    tx.commit().with_context(|| {
+        format!(
+            "commit schema v24 to v25 migration; backup {} remains",
+            backup.display()
+        )
+    })?;
+    Ok(())
+}
diff --git a/crates/host-core/src/db/tests.rs b/crates/host-core/src/db/tests.rs
index 8f5ed4f1ea..bcfb8e1ad8 100644
--- a/crates/host-core/src/db/tests.rs
+++ b/crates/host-core/src/db/tests.rs
@@ -294,6 +294,43 @@ fn worktree_registry_enforces_one_row_per_owner_and_unique_ports() {
 /// `sessions` stamps nor the `session_todo` table, so the upgrade must add
 /// both, keep the v20 rows, leave a readable pre-migration backup, and stay
 /// idempotent when a fixture downgrades the same file in place.
+/// v25 adds the per-session `/fast` column (ADR 0332). Existing sessions keep
+/// their configuration, default to the normal tier, and a backup remains.
+#[test]
+fn v24_database_gains_session_fast_mode_defaulting_off() {
+    let dir = tempfile::tempdir().unwrap();
+    let path = dir.path().join("pi.sqlite");
+    let session_id = {
+        let db = Database::open(&path).unwrap();
+        let session = crate::sessions::create_session_with_thinking(
+            &db,
+            Some("Kept".into()),
+            None,
+            None,
+            None,
+            None,
+            Some("high".into()),
+        )
+        .unwrap();
+        db.conn()
+            .execute_batch("ALTER TABLE sessions DROP COLUMN fast_mode;")
+            .unwrap();
+        db.conn().pragma_update(None, "user_version", 24).unwrap();
+        session.id
+    };
+
+    let db = Database::open(&path).unwrap();
+    assert_eq!(schema_version(db.conn()), SCHEMA_VERSION);
+    assert!(migration_backup_path(&path, 24).exists());
+    let summary = crate::sessions::get_session(&db, &session_id)
+        .unwrap()
+        .unwrap()
+        .summary;
+    assert_eq!(summary.title, "Kept");
+    assert_eq!(summary.thinking_level, "high");
+    assert!(!summary.fast_mode);
+}
+
 #[test]
 fn v20_database_migrates_the_session_checklist_with_a_backup() {
     let dir = tempfile::tempdir().unwrap();
diff --git a/crates/host-core/src/plugins/providers.rs b/crates/host-core/src/plugins/providers.rs
index 2fa9ed15d8..4f4f2469e5 100644
--- a/crates/host-core/src/plugins/providers.rs
+++ b/crates/host-core/src/plugins/providers.rs
@@ -186,6 +186,7 @@ pub(crate) fn declared_providers(manifest: &PluginManifest) -> Vec<DeclaredPlugi
                                 supports_documents: None,
                                 available_for_subagents: None,
                                 native_web_search: None,
+                                fast_mode_allowed: None,
                             })
                         })
                         .collect::<Vec<_>>()
diff --git a/crates/host-core/src/plugins/providers/tests.rs b/crates/host-core/src/plugins/providers/tests.rs
index bab05ca211..433aa876c6 100644
--- a/crates/host-core/src/plugins/providers/tests.rs
+++ b/crates/host-core/src/plugins/providers/tests.rs
@@ -99,10 +99,10 @@ fn a_new_database_carries_the_owner_column_at_the_current_schema_version() {
     let (_dir, db, _secrets) = test_context();
     // v17 added the owner column, v18 the turn-queue priority column,
     // v19 session omit, v21 the session Todo checklist, and v22 the
-    // session-list index, v23 the session title source, and v24 the worktree
-    // registry. A fresh database is stamped with the newest version, so the
-    // column set is the current one.
-    assert_eq!(SCHEMA_VERSION, 24);
+    // session-list index, v23 the session title source, v24 the worktree
+    // registry, and v25 the session Fast mode switch. A fresh database is
+    // stamped with the newest version, so the column set is the current one.
+    assert_eq!(SCHEMA_VERSION, 25);
     let version: i64 = db
         .conn()
         .query_row("PRAGMA user_version", [], |row| row.get(0))
diff --git a/crates/host-core/src/providers/catalog.rs b/crates/host-core/src/providers/catalog.rs
index 76472dfd60..2df72fe644 100644
--- a/crates/host-core/src/providers/catalog.rs
+++ b/crates/host-core/src/providers/catalog.rs
@@ -81,6 +81,7 @@ pub(crate) fn normalize_model_bindings(bindings: &[ModelBinding]) -> Vec<ModelBi
                 supports_documents: binding.supports_documents,
                 available_for_subagents: binding.available_for_subagents,
                 native_web_search: binding.native_web_search,
+                fast_mode_allowed: binding.fast_mode_allowed,
             })
         })
         .collect()
@@ -104,6 +105,7 @@ fn legacy_model_binding(model_id: Option<String>) -> Vec<ModelBinding> {
                 supports_documents: None,
                 available_for_subagents: None,
                 native_web_search: None,
+                fast_mode_allowed: None,
             }]
         })
         .unwrap_or_default()
@@ -490,6 +492,7 @@ mod tests {
             supports_documents: None,
             available_for_subagents: None,
             native_web_search: None,
+            fast_mode_allowed: None,
         }
     }
 
diff --git a/crates/host-core/src/providers/model.rs b/crates/host-core/src/providers/model.rs
index 40f590012a..fa980ae3f2 100644
--- a/crates/host-core/src/providers/model.rs
+++ b/crates/host-core/src/providers/model.rs
@@ -170,6 +170,10 @@ pub struct ModelBinding {
     /// default because models.dev does not publish hosted-tool capability.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub native_web_search: Option<bool>,
+    /// Opt-in that lets the session `/fast` switch request OpenAI's priority
+    /// service tier for this model (ADR 0332). None/false keeps the default.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub fast_mode_allowed: Option<bool>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
diff --git a/crates/host-core/src/providers/repository.rs b/crates/host-core/src/providers/repository.rs
index 7dc2ed4b32..6ca8a8e26e 100644
--- a/crates/host-core/src/providers/repository.rs
+++ b/crates/host-core/src/providers/repository.rs
@@ -111,6 +111,7 @@ fn hydrate_dynamic_plugin_models(db: &Database, provider: &mut ProviderPublic) -
                 supports_documents: None,
                 available_for_subagents: None,
                 native_web_search: None,
+                fast_mode_allowed: None,
             }
         })
         .collect();
diff --git a/crates/host-core/src/providers/tests.rs b/crates/host-core/src/providers/tests.rs
index 6e953e1fdf..660cbc461e 100644
--- a/crates/host-core/src/providers/tests.rs
+++ b/crates/host-core/src/providers/tests.rs
@@ -253,6 +253,7 @@ fn model_bindings_roundtrip_and_legacy_model_migrates_on_read() {
                     supports_documents: None,
                     available_for_subagents: Some(true),
                     native_web_search: None,
+                    fast_mode_allowed: Some(true),
                 },
                 ModelBinding {
                     id: "plain-model".into(),
@@ -268,6 +269,7 @@ fn model_bindings_roundtrip_and_legacy_model_migrates_on_read() {
                     supports_documents: Some(false),
                     available_for_subagents: None,
                     native_web_search: None,
+                    fast_mode_allowed: None,
                 },
             ]),
             default_model_id: None,
@@ -295,6 +297,8 @@ fn model_bindings_roundtrip_and_legacy_model_migrates_on_read() {
     assert_eq!(provider.models[1].default_thinking_level, None);
     assert_eq!(provider.models[0].available_for_subagents, Some(true));
     assert_eq!(provider.models[1].available_for_subagents, None);
+    assert_eq!(provider.models[0].fast_mode_allowed, Some(true));
+    assert_eq!(provider.models[1].fast_mode_allowed, None);
 
     let raw: String = db
         .conn()
@@ -318,6 +322,8 @@ fn model_bindings_roundtrip_and_legacy_model_migrates_on_read() {
     assert_eq!(config["models"][1]["supportsDocuments"], false);
     assert!(config["models"][1].get("availableForSubagents").is_none());
     assert!(config["models"][1].get("supportsImages").is_none());
+    assert_eq!(config["models"][0]["fastModeAllowed"], true);
+    assert!(config["models"][1].get("fastModeAllowed").is_none());
     assert_eq!(provider.models[0].supports_images, Some(true));
     assert_eq!(provider.models[1].supports_documents, Some(false));
 
@@ -377,6 +383,7 @@ fn binding_with_alias(id: &str, alias: Option<&str>) -> ModelBinding {
         supports_documents: None,
         available_for_subagents: None,
         native_web_search: None,
+        fast_mode_allowed: None,
     }
 }
 
@@ -1560,6 +1567,7 @@ fn binding_with_limits(id: &str, context_window: u32, max_tokens: u32) -> ModelB
         supports_documents: None,
         available_for_subagents: None,
         native_web_search: None,
+        fast_mode_allowed: None,
     }
 }
 
diff --git a/crates/host-core/src/rpc/mod.rs b/crates/host-core/src/rpc/mod.rs
index 4f86af66bd..10bb144b15 100644
--- a/crates/host-core/src/rpc/mod.rs
+++ b/crates/host-core/src/rpc/mod.rs
@@ -719,6 +719,20 @@ fn plugin_session_rpc_err(error: impl ToString) -> JsonRpcError {
     rpc_err(rpc_code, message, &code)
 }
 
+/// Optional session `/fast` switch (ADR 0332). Absent or null leaves the
+/// stored value; anything but a boolean is rejected rather than coerced.
+fn fast_mode_param(params: &Value) -> Result<Option<bool>, JsonRpcError> {
+    match params.get("fastMode") {
+        None | Some(Value::Null) => Ok(None),
+        Some(Value::Bool(value)) => Ok(Some(*value)),
+        Some(_) => Err(rpc_err(
+            1002,
+            "fastMode must be a boolean",
+            "INVALID_PARAMS",
+        )),
+    }
+}
+
 /// Parse the optional session thinking selector at the RPC boundary.  A
 /// missing/null value keeps the backwards-compatible default; present values
 /// must be strings from the host's allowlist rather than being silently
@@ -2555,6 +2569,7 @@ async fn handle_request(
         }
         "session.create" => {
             let thinking_level = thinking_level_param(&params)?;
+            let fast_mode = fast_mode_param(&params)?.unwrap_or(false);
             let permission_parent = permission_parent_param(&params)?;
             // ADR 0329: an isolated session gets its own git worktree.
             let isolated = params
@@ -2616,6 +2631,7 @@ async fn handle_request(
                         .map(str::to_string),
                     thinking_level,
                     permission_mode,
+                    fast_mode,
                 },
             )
             .map_err(|e| rpc_err(1000, e.to_string(), "INTERNAL"))?;
@@ -2782,8 +2798,9 @@ async fn handle_request(
                 .and_then(|v| v.as_str())
                 .ok_or_else(|| rpc_err(1002, "mode required", "INVALID_PARAMS"))?;
             let thinking_level = thinking_level_param(&params)?;
+            let fast_mode = fast_mode_param(&params)?;
             let st = state.lock().await;
-            let session = sessions::configure_session_with_thinking(
+            let session = sessions::configure_session_with_fast_mode(
                 &st.db,
                 id,
                 mode,
@@ -2791,6 +2808,7 @@ async fn handle_request(
                 params.get("modelId").and_then(|v| v.as_str()),
                 thinking_level.as_deref(),
                 params.get("permissionMode").and_then(|v| v.as_str()),
+                fast_mode,
             )
             .map_err(|e| {
                 let message = e.to_string();
@@ -6682,6 +6700,58 @@ mod tests {
         );
     }
 
+    /// ADR 0332: `fastMode` is an optional boolean on create and configure;
+    /// an absent value keeps the stored switch and a non-boolean is rejected.
+    #[tokio::test]
+    async fn session_fast_mode_rpc_is_optional_boolean_and_preserved() {
+        let data_dir = tempfile::tempdir().unwrap();
+        let mut app_state = AppState::open(data_dir.path()).unwrap();
+        app_state.handshook = true;
+        let state = Arc::new(Mutex::new(app_state));
+        let (tx, _rx) = mpsc::unbounded_channel();
+        let call = |method: &'static str, params: Value| {
+            handle_request(state.clone(), method, params, tx.clone())
+        };
+
+        let legacy = call("session.create", json!({ "mode": "agent" }))
+            .await
+            .unwrap();
+        assert_eq!(legacy["session"]["fastMode"], false);
+        let created = call(
+            "session.create",
+            json!({ "mode": "agent", "fastMode": true }),
+        )
+        .await
+        .unwrap();
+        assert_eq!(created["session"]["fastMode"], true);
+        let id = created["session"]["id"].as_str().unwrap().to_string();
+
+        let kept = call(
+            "session.configure",
+            json!({ "id": id, "mode": "plan", "thinkingLevel": "high" }),
+        )
+        .await
+        .unwrap();
+        assert_eq!(kept["session"]["fastMode"], true);
+        let off = call(
+            "session.configure",
+            json!({ "id": id, "mode": "agent", "fastMode": false }),
+        )
+        .await
+        .unwrap();
+        assert_eq!(off["session"]["fastMode"], false);
+
+        for method in ["session.create", "session.configure"] {
+            let error = call(
+                method,
+                json!({ "id": id, "mode": "agent", "fastMode": "yes" }),
+            )
+            .await
+            .unwrap_err();
+            assert_eq!(error.data.unwrap()["errorCode"], "INVALID_PARAMS");
+        }
+    }
+
     #[tokio::test]
     async fn session_create_rpc_inherits_optional_permission_mode() {
         let data_dir = tempfile::tempdir().unwrap();
diff --git a/crates/host-core/src/rpc/scheduled_rpc.rs b/crates/host-core/src/rpc/scheduled_rpc.rs
index 7e043f47ab..f86ca943ac 100644
--- a/crates/host-core/src/rpc/scheduled_rpc.rs
+++ b/crates/host-core/src/rpc/scheduled_rpc.rs
@@ -192,6 +192,8 @@ fn handle_with_workspace_policy(
                     .permission_mode
                     .clone()
                     .or_else(|| automatic.then(|| "ask".into())),
+                // Unattended runs keep the default tier (ADR 0332).
+                fast_mode: false,
             };
             let uses_task_execution_settings = task.permission_mode.is_some()
                 || task.thinking_level.is_some()
diff --git a/crates/host-core/src/session_collaboration/mod.rs b/crates/host-core/src/session_collaboration/mod.rs
index 50077ead61..dac851691b 100644
--- a/crates/host-core/src/session_collaboration/mod.rs
+++ b/crates/host-core/src/session_collaboration/mod.rs
@@ -167,6 +167,8 @@ fn spawn(db: &Database, input: &Value) -> Result<Value> {
             project_path: parent.summary.project_path,
             thinking_level: Some(parent.summary.thinking_level),
             permission_mode: Some(parent.summary.permission_mode),
+            // Delegated workers keep the default tier (ADR 0332).
+            fast_mode: false,
         },
     )?;
     db.conn().execute("INSERT INTO session_collaboration_links(session_id,created_by_session_id,plugin_id,created_at) VALUES(?1,?2,?3,?4)",params![created.id,source,plugin,now_ms()])?;
diff --git a/crates/host-core/src/session_search.rs b/crates/host-core/src/session_search.rs
index f291265f10..a1aac3775a 100644
--- a/crates/host-core/src/session_search.rs
+++ b/crates/host-core/src/session_search.rs
@@ -107,7 +107,8 @@ pub fn search(db: &Database, query: &str, offset: i64, limit: i64) -> Result<Sea
                 p.name, COALESCE(matched.count, 0),
                 (pi_search_contains(s.title, ?1) OR pi_search_contains(p.name, ?1)
                  OR pi_search_contains(p.path, ?1)) AS metadata_match,
-                EXISTS (SELECT 1 FROM task_runs r WHERE r.session_id = s.id) AS scheduled_run
+                EXISTS (SELECT 1 FROM task_runs r WHERE r.session_id = s.id) AS scheduled_run,
+                s.fast_mode
          FROM sessions s LEFT JOIN projects p ON p.id = s.project_id
          LEFT JOIN matched ON matched.session_id = s.id
          WHERE s.deleted_at IS NULL AND (matched.count > 0 OR metadata_match)
diff --git a/crates/host-core/src/session_worktrees/delegate/tests.rs b/crates/host-core/src/session_worktrees/delegate/tests.rs
index 88c5e8de0c..f66ab539f2 100644
--- a/crates/host-core/src/session_worktrees/delegate/tests.rs
+++ b/crates/host-core/src/session_worktrees/delegate/tests.rs
@@ -40,6 +40,7 @@ async fn env() -> Env {
                 project_path: Some(project.to_string_lossy().into_owned()),
                 thinking_level: None,
                 permission_mode: None,
+                fast_mode: false,
             },
         )
         .unwrap()
diff --git a/crates/host-core/src/session_worktrees/lifecycle/tests.rs b/crates/host-core/src/session_worktrees/lifecycle/tests.rs
index 708d6855c5..6868ba08f0 100644
--- a/crates/host-core/src/session_worktrees/lifecycle/tests.rs
+++ b/crates/host-core/src/session_worktrees/lifecycle/tests.rs
@@ -41,6 +41,7 @@ async fn new_session(env: &Env, project: &Path) -> String {
             project_path: Some(project.to_string_lossy().into_owned()),
             thinking_level: None,
             permission_mode: None,
+            fast_mode: false,
         },
     )
     .unwrap()
@@ -259,6 +260,7 @@ fn new_session_blocking(db: &Database, project: &Path) -> String {
             project_path: Some(project.to_string_lossy().into_owned()),
             thinking_level: None,
             permission_mode: None,
+            fast_mode: false,
         },
     )
     .unwrap()
diff --git a/crates/host-core/src/sessions.rs b/crates/host-core/src/sessions.rs
index df89a3dcc5..57f63bc551 100644
--- a/crates/host-core/src/sessions.rs
+++ b/crates/host-core/src/sessions.rs
@@ -128,6 +128,10 @@ pub struct SessionSummary {
     pub thinking_level: String,
     #[serde(default = "default_permission_mode")]
     pub permission_mode: String,
+    /// The `/fast` switch for this session's main turns (ADR 0332). Requests
+    /// use the priority tier only when the model binding also allows it.
+    #[serde(default)]
+    pub fast_mode: bool,
     pub updated_at: String,
     pub created_at: String,
     /// True when this session is the transcript owned by a scheduled-task run.
@@ -1290,7 +1294,8 @@ fn session_created_at(db: &Database, session_id: &str) -> Result<String> {
 const SUMMARY_SELECT: &str =
     "SELECT s.id, s.title, s.last_seq, p.path, s.model_id, s.provider_id, s.mode,
             s.thinking_level, s.permission_mode, s.updated_at, s.created_at,
-            EXISTS (SELECT 1 FROM task_runs r WHERE r.session_id = s.id) AS scheduled_run
+            EXISTS (SELECT 1 FROM task_runs r WHERE r.session_id = s.id) AS scheduled_run,
+            s.fast_mode
      FROM sessions s LEFT JOIN projects p ON p.id = s.project_id
      WHERE s.deleted_at IS NULL";
 
@@ -1310,6 +1315,7 @@ pub(crate) fn summary_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Sess
         // Read by name: search and listing build their own column lists, so the
         // alias keeps this mapper independent of any one query's column order.
         scheduled_run: row.get("scheduled_run")?,
+        fast_mode: row.get("fast_mode")?,
         worktree: None,
     })
 }
@@ -1365,6 +1371,8 @@ pub struct SessionCreateOptions {
     pub project_path: Option<String>,
     pub thinking_level: Option<String>,
     pub permission_mode: Option<String>,
+    /// The `/fast` switch (ADR 0332); false keeps the default service tier.
+    pub fast_mode: bool,
 }
 
 pub fn create_session_with_thinking(
@@ -1386,6 +1394,7 @@ pub fn create_session_with_thinking(
             project_path,
             thinking_level,
             permission_mode: None,
+            fast_mode: false,
         },
     )
 }
@@ -1407,6 +1416,7 @@ pub fn create_session_with_options(
         project_path,
         thinking_level,
         permission_mode,
+        fast_mode,
     } = options;
     let now = now_ms();
     let id = Uuid::new_v4().to_string();
@@ -1436,8 +1446,8 @@ pub fn create_session_with_options(
         .prepare_cached(
             "INSERT INTO sessions (
                 id, title, project_id, provider_id, model_id, mode, thinking_level,
-                permission_mode, title_source, created_at, updated_at
-             ) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?10)",
+                permission_mode, title_source, fast_mode, created_at, updated_at
+             ) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?11, ?11)",
         )?
         .execute(params![
             id,
@@ -1449,6 +1459,7 @@ pub fn create_session_with_options(
             thinking_level,
             permission_mode,
             title_source,
+            fast_mode,
             now
         ])?;
     Ok(SessionSummary {
@@ -1461,6 +1472,7 @@ pub fn create_session_with_options(
         mode,
         thinking_level,
         permission_mode,
+        fast_mode,
         updated_at: ms_to_ts(now),
         created_at: ms_to_ts(now),
         // The run row that owns this transcript is written after creation.
@@ -1793,10 +1805,11 @@ pub fn fork_session_through(
             .prepare_cached(
                 "INSERT INTO sessions (
                     id, title, project_id, provider_id, model_id, mode, thinking_level,
-                    permission_mode, source, pinned, last_seq, created_at, updated_at
+                    permission_mode, fast_mode, source, pinned, last_seq, created_at,
+                    updated_at
                  )
                  SELECT ?1, ?2, project_id, provider_id, model_id, mode, thinking_level,
-                        permission_mode, NULL, 0, ?3, ?4, ?4
+                        permission_mode, fast_mode, NULL, 0, ?3, ?4, ?4
                  FROM sessions WHERE id = ?5",
             )?
             .execute(params![id, title, records.len() as i64, now, source_id])?;
@@ -1825,6 +1838,7 @@ pub fn fork_session_through(
         mode: source.summary.mode,
         thinking_level: source.summary.thinking_level,
         permission_mode: source.summary.permission_mode,
+        fast_mode: source.summary.fast_mode,
         updated_at: created_at.clone(),
         created_at,
         // A fork is the user's own conversation, not the automation's transcript.
@@ -1869,6 +1883,31 @@ pub fn configure_session_with_thinking(
     model_id: Option<&str>,
     thinking_level: Option<&str>,
     permission_mode: Option<&str>,
+) -> Result<Option<SessionSummary>> {
+    configure_session_with_fast_mode(
+        db,
+        id,
+        mode,
+        provider_id,
+        model_id,
+        thinking_level,
+        permission_mode,
+        None,
+    )
+}
+
+/// Full configurator. Every `None` preserves the persisted value, including
+/// the `/fast` switch (ADR 0332).
+#[allow(clippy::too_many_arguments)]
+pub fn configure_session_with_fast_mode(
+    db: &Database,
+    id: &str,
+    mode: &str,
+    provider_id: Option<&str>,
+    model_id: Option<&str>,
+    thinking_level: Option<&str>,
+    permission_mode: Option<&str>,
+    fast_mode: Option<bool>,
 ) -> Result<Option<SessionSummary>> {
     if !(is_valid_mode(mode) || mode == "chat") {
         return Err(anyhow!("mode must be plan or agent"));
@@ -1896,7 +1935,8 @@ pub fn configure_session_with_thinking(
              SET mode = ?2, provider_id = COALESCE(?3, provider_id),
                  model_id = COALESCE(?4, model_id),
                  thinking_level = COALESCE(?5, thinking_level),
-                 permission_mode = COALESCE(?6, permission_mode), updated_at = ?7
+                 permission_mode = COALESCE(?6, permission_mode),
+                 fast_mode = COALESCE(?8, fast_mode), updated_at = ?7
              WHERE id = ?1",
         )?
         .execute(params![
@@ -1906,7 +1946,8 @@ pub fn configure_session_with_thinking(
             model_id,
             thinking_level,
             permission_mode,
-            now_ms()
+            now_ms(),
+            fast_mode
         ])?;
     if changed == 0 {
         return Ok(None);
@@ -4746,6 +4787,76 @@ mod tests {
         assert_eq!(omitted.thinking_level, "omit");
     }
 
+    /// ADR 0332: `/fast` is stored per session, survives configures that do
+    /// not mention it, and is copied by a fork.
+    #[test]
+    fn session_fast_mode_is_persisted_preserved_and_forked() {
+        let db = test_db();
+        let session = create_session(&db, None, None, None, None, None).unwrap();
+        assert!(!session.fast_mode);
+
+        let set = |enabled: Option<bool>, mode: &str| {
+            configure_session_with_fast_mode(
+                &db,
+                &session.id,
+                mode,
+                None,
+                None,
+                None,
+                None,
+                enabled,
+            )
+            .unwrap()
+            .unwrap()
+        };
+        assert!(set(Some(true), "agent").fast_mode);
+        assert!(set(None, "plan").fast_mode);
+        let kept = configure_session(&db, &session.id, "agent", None, None)
+            .unwrap()
+            .unwrap();
+        assert!(kept.fast_mode);
+        assert!(
+            get_session(&db, &session.id)
+                .unwrap()
+                .unwrap()
+                .summary
+                .fast_mode
+        );
+
+        let ForkSessionResult::Created(fork) =
+            fork_session_through(&db, &session.id, None, None).unwrap()
+        else {
+            panic!("expected fork");
+        };
+        assert!(fork.summary.fast_mode);
+        assert!(
+            get_session(&db, &fork.summary.id)
+                .unwrap()
+                .unwrap()
+                .summary
+                .fast_mode
+        );
+
+        assert!(!set(Some(false), "agent").fast_mode);
+
+        let created = create_session_with_options(
+            &db,
+            SessionCreateOptions {
+                fast_mode: true,
+                ..Default::default()
+            },
+        )
+        .unwrap();
+        assert!(created.fast_mode);
+        assert!(
+            get_session(&db, &created.id)
+                .unwrap()
+                .unwrap()
+                .summary
+                .fast_mode
+        );
+    }
+
     #[test]
     fn create_session_permission_mode_is_optional_and_validated() {
         let db = test_db();
@@ -5038,6 +5149,7 @@ mod tests {
             permission_mode: "inherit".into(),
             // Ownership is derived from `task_runs`, so an import is never one.
             scheduled_run: false,
+            fast_mode: false,
             worktree: None,
             created_at: "2025-01-01T00:00:00Z".into(),
             updated_at: "2025-01-02T00:00:00Z".into(),
@@ -5091,6 +5203,7 @@ mod tests {
             permission_mode: "inherit".into(),
             // Ownership is derived from `task_runs`, so an import is never one.
             scheduled_run: false,
+            fast_mode: false,
             worktree: None,
             created_at: "2025-01-01T00:00:00Z".into(),
             updated_at: "2025-01-01T00:00:00Z".into(),
@@ -6033,6 +6146,7 @@ mod tests {
             permission_mode: "inherit".into(),
             // Ownership is derived from `task_runs`, so an import is never one.
             scheduled_run: false,
+            fast_mode: false,
             worktree: None,
             created_at: "2025-01-01T00:00:00Z".into(),
             updated_at: "2025-01-01T00:00:00Z".into(),
diff --git a/crates/host-core/src/turn_queue.rs b/crates/host-core/src/turn_queue.rs
index 92e9b37334..6ca93814e3 100644
--- a/crates/host-core/src/turn_queue.rs
+++ b/crates/host-core/src/turn_queue.rs
@@ -532,7 +532,7 @@ mod tests {
         assert_eq!(entries[0].content, "existing");
         assert!(entries[0].user_message_id.is_none());
         assert!(entries[0].voice_origin.is_none());
-        assert_eq!(crate::db::SCHEMA_VERSION, 24);
+        assert_eq!(crate::db::SCHEMA_VERSION, 25);
     }
 
     #[test]
diff --git a/docs/adr/0332-openai-fast-mode.md b/docs/adr/0332-openai-fast-mode.md
new file mode 100644
index 0000000000..1c1a47cec2
--- /dev/null
+++ b/docs/adr/0332-openai-fast-mode.md
@@ -0,0 +1,69 @@
+# ADR 0332: OpenAI Fast mode as a per-session switch gated by model settings
+
+- Status: Accepted
+- Date: 2026-10-17
+- Deciders: PI-Desktop maintainers
+- Related: [ADR 0329](0329-session-worktree-isolation.md) ·
+  [03-runtime/11-provider-model-system](../spec/03-runtime/11-provider-model-system.md) ·
+  [04-ux/04-builtin-commands](../spec/04-ux/04-builtin-commands.md) ·
+  issue #1657 · E2E-FAST-MODE
+
+## Context
+
+OpenAI serves a faster priority tier that Codex CLI and the OpenAI API call
+Fast mode. On the wire it is the Responses body field `service_tier:
+"priority"`; the Codex backend rejects `"fast"`. The tier costs more: a ChatGPT
+subscription spends its plan limits faster, and the OpenAI API bills a higher
+per-token price. PI-Desktop never asked for a tier, and pi-ai's `streamSimple`
+does not forward one.
+
+The cost is a decision about one conversation, while whether an endpoint
+honors the field is something only the user knows for a relay. Issue #1657 asks
+for the subscription speed-up; maintainers asked that capability be configured
+in model settings.
+
+## Decision
+
+1. **Model settings own capability.** `ModelBinding.fastModeAllowed` is an
+   opt-in, stored in the existing provider `config_json` like
+   `nativeWebSearch`, so it needs no migration. The checkbox is enabled only
+   for Responses wires (`openai_codex_responses`, `responses`), including
+   relays; its hint names the cost on that wire.
+2. **`/fast` owns the session switch.** A new builtin `builtin.session.fast`
+   (`/fast`) toggles a persisted `sessions.fast_mode` column (schema v25,
+   additive, v24 backup). `session.create` and `session.configure` accept an
+   optional boolean `fastMode`; omission keeps the stored value. Forks copy it;
+   scheduled runs and delegated workers start off. A draft carries it into the
+   session its first prompt creates. Remote sessions refuse it.
+3. **`/fast` only works on a configured model.** When the selected model's
+   wire cannot carry the tier or the model is not allowed, `/fast` stores
+   nothing and points to Settings → Models → Advanced, in either direction. A
+   stale stored switch on such a model is inert because launch also requires
+   the opt-in.
+4. **The launcher resolves; the runtime only stamps the wire.** Session launch
+   sends `fastMode: true` to the sidecar only when the session switch is on and
+   the user's stored binding allows it. The runtime chains a payload hook on
+   main turns that adds `service_tier: "priority"` when the resolved wire is
+   `openai-codex-responses` or `openai-responses`. Subagents, compaction, and
+   one-shot plugin requests build their own options and keep the default tier.
+
+## Alternatives
+
+- **Per-model switch only (no session state).** Avoids a migration, but turns
+  a per-conversation cost into a global one: enabling Fast for one task speeds
+  up and costs every conversation on that model.
+- **Hard-coded capability rule (official OpenAI host only).** Removes a
+  setting, but excludes relays the user knows support the field and hides the
+  decision in code.
+- **Generic request-parameter injection with presets.** More flexible, but an
+  arbitrary body-field channel needs its own validation and permission story,
+  which is larger than this need.
+
+## Consequences
+
+- Schema v25 adds one column; older apps opening a v25 database refuse it as
+  newer, the existing downgrade behavior.
+- The capability can be wrong for a relay that ignores or rebills the field;
+  the settings hint states that risk, and the user owns the opt-in.
+- The Composer model chip shows a Fast mark while the switch is on for an
+  allowed model, so the cost is visible at send time.
diff --git a/docs/adr/README.md b/docs/adr/README.md
index 728f4e585c..922f4a80c4 100644
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -377,6 +377,7 @@ Each ADR includes:
 | 0329 | [Isolate parallel sessions of one project in git worktrees](0329-session-worktree-isolation.md) | Accepted (Windows E2E pending) |
 | 0330 | [Open scoped HTML files in the work-panel Browser from plugins](0330-open-html-files-in-work-panel-browser-from-plugins.md) | Accepted |
 | 0331 | [Plugin session follow and host-owned work-panel reveal](0331-plugin-session-follow-and-work-panel-reveal.md) | Accepted |
+| 0332 | [OpenAI Fast mode as a per-session switch gated by model settings](0332-openai-fast-mode.md) | Accepted |
 | plugin-fetch-redirect-policy | [Host-enforced plugin fetch redirect policy](plugin-fetch-redirect-policy.md) | Accepted (implementation candidate for #1475) |
 
 | local-agent-remote-workspace-tools | [Local Agent and model requests for remote workspaces](local-agent-remote-workspace-tools.md) | Accepted for implementation |
diff --git a/docs/project/unreleased.md b/docs/project/unreleased.md
index 367d666642..69b5047677 100644
--- a/docs/project/unreleased.md
+++ b/docs/project/unreleased.md
@@ -1 +1,7 @@
 # Unreleased changes
+
+- Add OpenAI Fast mode (#1657). Allow it per model in Settings → Models for a
+  ChatGPT subscription or an OpenAI Responses connection, then type `/fast` to
+  turn it on for the current conversation. Requests ask for the priority
+  service tier, which replies faster and spends plan limits faster or bills
+  more per token; the Composer model chip shows a Fast mark while it is on.
diff --git a/docs/spec/03-runtime/01-ipc-protocol.md b/docs/spec/03-runtime/01-ipc-protocol.md
index 64c0fa9051..29d19104f7 100644
--- a/docs/spec/03-runtime/01-ipc-protocol.md
+++ b/docs/spec/03-runtime/01-ipc-protocol.md
@@ -188,6 +188,8 @@ type SessionConfigureRequest = {
   providerId?: string;
   modelId?: string;
   thinkingLevel: SessionThinkingLevel;
+  /** Session `/fast` switch (ADR 0332); omitted keeps the stored value. */
+  fastMode?: boolean;
 };
 ```
 
diff --git a/docs/spec/03-runtime/04-data-storage.md b/docs/spec/03-runtime/04-data-storage.md
index df5d9d2aae..b0b740862a 100644
--- a/docs/spec/03-runtime/04-data-storage.md
+++ b/docs/spec/03-runtime/04-data-storage.md
@@ -1,4 +1,4 @@
-# 04. Data Storage (Schema v23)
+# 04. Data Storage (Schema v25)
 
 ## 0. Ownership decision
 
@@ -585,6 +585,8 @@ CREATE TABLE sessions (
                                           'high', 'xhigh', 'max', 'omit')),
   permission_mode TEXT NOT NULL DEFAULT 'inherit' -- D115: inherit follows settings
                 CHECK (permission_mode IN ('inherit', 'ask', 'accept-edits', 'auto')),
+  fast_mode   INTEGER NOT NULL DEFAULT 0       -- ADR 0332: session /fast switch
+               CHECK (fast_mode IN (0, 1)),
   source      TEXT,                            -- import origin: claude-code | codex | opencode | pi
   deleted_at  INTEGER,                         -- plugin trash marker; null means active
   pinned      INTEGER NOT NULL DEFAULT 0,
@@ -1544,7 +1546,7 @@ truncating at a guessed position.
 - JSON columns are read blind on hot paths (shipped to the renderer as-is);
   anything filtered or summed is a promoted column by rule.
 
-## 7. Versioning, v7 reset, and v8-to-v23 migration
+## 7. Versioning, v7 reset, and v8-to-v25 migration
 
 - `PRAGMA user_version` stays the schema authority; future structural changes
   add ordered Rust migration fns again, each in one transaction, with a
@@ -1555,7 +1557,7 @@ truncating at a guessed position.
   Sessions, providers, and settings from the old file are not carried over;
   the archive remains for manual recovery. All pre-v7 migration code
   (v1 `settings.sqlite` import, v2→v6 chain) is deleted.
-- Fresh installs run the full v23 DDL directly.
+- Fresh installs run the full v25 DDL directly.
 - **Schema v7 first reaches v8, then uses the guarded path.** The v7→v8
   migration is followed by the same guarded v8→v15 migration; schema-v9 and
   schema-v10 databases take the same guarded path and receive an exact readable
@@ -1619,6 +1621,13 @@ truncating at a guessed position.
 - **Schema v23 is additive.** It adds `sessions.title_source` and classifies
   existing placeholder titles as `default`; every other existing title is
   preserved and classified as `manual`. The migration keeps a v22 backup.
+- **Schema v24 is additive.** It adds the `worktrees` registry and
+  `projects.worktree_setup_command` for isolated sessions (ADR 0329). The
+  migration keeps a v23 backup.
+- **Schema v25 is additive.** It adds `sessions.fast_mode` (ADR 0332), the
+  session `/fast` switch, defaulting to `0` for every existing session. Forks
+  copy it; scheduled runs and delegated workers start at `0`. The migration
+  keeps a v24 backup.
 - **Schema v14 is additive.** It adds nullable `sessions.deleted_at`, the
   partial deletion index, and `session_import_origins`. Existing sessions stay
   active and have no origin rows. The migration runs in the same guarded
diff --git a/docs/spec/03-runtime/06-host-rpc-protocol.md b/docs/spec/03-runtime/06-host-rpc-protocol.md
index e03cfbd781..c103419bb5 100644
--- a/docs/spec/03-runtime/06-host-rpc-protocol.md
+++ b/docs/spec/03-runtime/06-host-rpc-protocol.md
@@ -296,6 +296,8 @@ to later refresh and inference; the vendor picker does not collect them.
   `inheritPermissionFromSessionId`; when present, the host copies the existing
   session's persisted permission mode atomically, while omission preserves the
   existing `inherit` default. Missing/null thinking level defaults to `off`.
+  Optional boolean `fastMode` (ADR 0332) stores the session `/fast` switch;
+  missing/null means `false` and a non-boolean returns `INVALID_PARAMS`.
   Optional `isolated: true` (ADR 0329) also requires `projectPath` and gives
   the new session its own git worktree before it returns; if the worktree
   cannot be created the session is deleted again and the call fails with a
@@ -352,7 +354,9 @@ to later refresh and inference; the vendor picker does not collect them.
 - `session.configure` — atomically persists `mode`, `providerId`, `modelId`,
   and optional `thinkingLevel` (`off|minimal|low|medium|high|xhigh|max|omit`)
   for the next pi turn; omitting/null
-  `thinkingLevel` preserves the current value; invalid modes or levels return
+  `thinkingLevel` preserves the current value; optional boolean `fastMode`
+  (ADR 0332) follows the same rule, and a non-boolean is `INVALID_PARAMS`;
+  invalid modes or levels return
   `INVALID_PARAMS`; mode is `plan | goal | agent` and changing any session
   configuration is allowed only while idle and without a pending/queued/running
   Plan or Goal record
diff --git a/docs/spec/03-runtime/11-provider-model-system.md b/docs/spec/03-runtime/11-provider-model-system.md
index 30f0f76f88..dbf505947f 100644
--- a/docs/spec/03-runtime/11-provider-model-system.md
+++ b/docs/spec/03-runtime/11-provider-model-system.md
@@ -351,6 +351,7 @@ type ModelBinding = {
   defaultThinkingLevel: SessionThinkingLevel | null
   thinkingProtocol?: "legacy" | "adaptive"
   availableForSubagents?: boolean // opt-in for AI-driven delegation
+  fastModeAllowed?: boolean // opt-in for the session /fast switch (ADR 0332)
 }
 
 type SelectedModelRef = {
@@ -383,6 +384,17 @@ array remain future work. A legacy provider with only `defaultModelId` is
 materialized as one fallback binding on host read and upgraded to `models` on
 the next provider write.
 
+`ModelBinding.fastModeAllowed` (boolean, default false, ADR 0332): lets the
+session `/fast` switch request OpenAI's priority service tier
+(`service_tier: "priority"`) for this model. Model settings enable the checkbox
+only for Responses wires: `openai_codex_responses` (ChatGPT subscription; the
+tier spends plan limits faster) and `responses` (billed per token at the tier's
+higher price; a relay may ignore or rebill the field). The launcher sends
+`fastMode: true` to the sidecar only when the session switch is on and the
+stored binding allows it; the runtime adds the field to main-turn requests whose
+resolved wire is `openai-codex-responses` or `openai-responses`. Subagents,
+compaction, and one-shot requests keep the default tier.
+
 `ModelBinding.availableForSubagents` (boolean, default false): opt-in flag that
 makes the model available for AI-driven subagent delegation. When enabled, the
 model appears in the delegation catalog injected into the parent agent's system
diff --git a/docs/spec/04-ux/04-builtin-commands.md b/docs/spec/04-ux/04-builtin-commands.md
index 3b00e5779c..05c462326f 100644
--- a/docs/spec/04-ux/04-builtin-commands.md
+++ b/docs/spec/04-ux/04-builtin-commands.md
@@ -2,7 +2,7 @@
 
 ## 1. Goal
 
-Define the five first-party command palette entries available without plugins.
+Define the six first-party command palette entries available without plugins.
 The same entries are available in the composer `/` menu.
 
 Shortcut: **Cmd/Ctrl + Shift + P** (D014)
@@ -15,7 +15,7 @@ builtin.<domain>.<action>
 
 ## 3. Core builtin catalog
 
-The builtin registry intentionally contains exactly these five commands. Plugin
+The builtin registry intentionally contains exactly these six commands. Plugin
 commands extend the searchable command list dynamically; app navigation,
 project management, settings, plugin management, and diagnostics are not
 builtin command contracts.
@@ -27,10 +27,11 @@ builtin command contracts.
 | `builtin.mode.agent` | Switch to Agent | mode, agent | Session | low | set the idle session mode to Agent |
 | `builtin.mode.plan` | Switch to Plan | mode, plan, planning | Session | low | set the idle session mode to Plan |
 | `builtin.mode.goal` | Switch to Goal | mode, goal, objective, autonomous | Session | low | set the idle session mode to Goal |
+| `builtin.session.fast` | Toggle Fast mode | fast, speed, priority, service tier | Session | medium | toggle the session's OpenAI priority service tier (ADR 0332) |
 
 ## 4. Visibility and execution rules
 
-- The five IDs are the complete first-party registry. Removed IDs are not
+- The six IDs are the complete first-party registry. Removed IDs are not
   palette results and are not renderer dispatch cases; plugin commands remain
   independently discoverable.
 - `New Task` uses the current project or temporary group. It selects the
@@ -44,6 +45,15 @@ builtin command contracts.
   Agent/Plan/Goal chip. They update an idle session immediately. With no active
   session, they update the persisted default for the next session; a running
   session or pending approval is not changed.
+- `Toggle Fast mode` flips the session's persisted `fastMode` through the same
+  configuration path (ADR 0332). With no active session it records the switch on
+  the new-task draft, and the created session starts with it. The selected
+  model must allow Fast mode in Settings → Models → Advanced on a Responses
+  wire (Codex subscription or OpenAI Responses); otherwise, in either
+  direction, a warning toast points to that setting and nothing is stored. A
+  stored switch on an unconfigured model has no effect at launch. Remote
+  sessions refuse it. The risk is medium because the priority tier
+  spends plan limits faster or bills more per token.
 - `SubmitPlan` and `SubmitGoal` are model tools, not palette commands. There is
   no Chat mode or request-changes alias.
 - The former app/project/settings/plugin/log commands remain available through
@@ -62,7 +72,7 @@ type CommandExecutionResult =
 
 ## 6. Acceptance
 
-1. The builtin registry contains exactly five unique, prefixed IDs.
+1. The builtin registry contains exactly six unique, prefixed IDs.
 2. Palette search matches each title and keyword set.
 3. Mode switch commands update the idle session mode immediately; Plan, Goal,
    and Agent refer to the same pi Agent.
@@ -84,11 +94,12 @@ in the same registry that feeds palette search
 | `/agent-mode` | `builtin.mode.agent` |
 | `/plan-mode` | `builtin.mode.plan` |
 | `/goal-mode` | `builtin.mode.goal` |
+| `/fast` | `builtin.session.fast` |
 
 Aliases share one namespace with template and plugin command names; builtin
 aliases win collisions, then project templates, then user templates, then
-plugin commands. Selecting an alias inserts `/alias `; sending `/new` or
-`/compact` alone executes locally without creating an empty prompt. The
+plugin commands. Selecting an alias inserts `/alias `; sending `/new`,
+`/compact`, or `/fast` alone executes locally without creating an empty prompt. The
 Agent/Plan/Goal aliases also support a prompt body:
 `/agent-mode <prompt>`, `/plan-mode <prompt>`, or `/goal-mode <prompt>` switches
 the idle session (or the next-session default) and sends `<prompt>` through the
diff --git a/docs/spec/06-delivery/04-e2e-test-plan.md b/docs/spec/06-delivery/04-e2e-test-plan.md
index fc3b8c2607..a5192c1758 100644
--- a/docs/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/spec/06-delivery/04-e2e-test-plan.md
@@ -18217,3 +18217,34 @@ host-created files. The full app's file-preview viewer is covered separately.
   `ui.view` gating, panel-bridge channel) and the wiring contract in
   `apps/desktop/test/main-wiring-contract.test.mjs`; the interactive Electron
   scenario remains a manual validation run.
+
+### E2E-FAST-MODE: `/fast` asks for the priority tier only on an allowed model
+
+- **Preconditions**: A ChatGPT subscription (Codex) account and a Responses
+  API-key provider, each with two configured models. No real provider is
+  required for the automated path; the manual path uses a capturing local
+  endpoint or the user's authorized account.
+- **Steps**:
+  1. In Settings → Models, check **Fast mode** on one model of each provider.
+     Confirm the checkbox is disabled on a Chat Completions provider.
+  2. Start a new task on the allowed Codex model, type `/fast`, send a prompt.
+  3. Send `/fast` again, then send another prompt.
+  4. Switch the session to the model without the opt-in and send `/fast`.
+  5. Fork the session while Fast is on; restart the app and reopen both.
+- **Expected**: Step 1 persists `fastModeAllowed` on the binding only. Step 2
+  shows the Fast mark on the model chip and the request body carries
+  `service_tier: "priority"`; a Responses API-key model shows the per-token
+  price toast instead. Step 3 clears the mark and the field. Step 4 refuses
+  with the model-settings hint and stores nothing. Step 5 keeps the switch on
+  the fork and across restart. Subagent and compaction requests never carry
+  the field. Remote sessions refuse `/fast`.
+- **Specs linked**: `03-runtime/11-provider-model-system.md`,
+  `03-runtime/04-data-storage.md` (v25), `04-ux/04-builtin-commands.md`,
+  ADR 0332, issue #1657.
+- **Status**: Automated coverage: `apps/desktop/test/fast-mode-command.test.mjs`
+  (draft → created session, toggle, refusals, remote, chip),
+  `apps/desktop/test/fast-mode-session-launch.test.mjs` (launch resolution),
+  `packages/agent-runtime/src/fast-mode.test.ts` (request body through the real
+  pi-ai Responses and Completions adapters), host-core migration, RPC and fork
+  tests. The full settings → `/fast` → captured request path is a manual
+  validation scenario.
diff --git a/docs/zh-CN/spec/03-runtime/01-ipc-protocol.md b/docs/zh-CN/spec/03-runtime/01-ipc-protocol.md
index 79a7716a22..b3a25113ae 100644
--- a/docs/zh-CN/spec/03-runtime/01-ipc-protocol.md
+++ b/docs/zh-CN/spec/03-runtime/01-ipc-protocol.md
@@ -158,6 +158,8 @@ type SessionConfigureRequest = {
   providerId?: string;
   modelId?: string;
   thinkingLevel: SessionThinkingLevel;
+  /** Session `/fast` switch (ADR 0332); omitted keeps the stored value. */
+  fastMode?: boolean;
 };
 ```
 
diff --git a/docs/zh-CN/spec/03-runtime/04-data-storage.md b/docs/zh-CN/spec/03-runtime/04-data-storage.md
index 5b8bc49f74..33c9c17115 100644
--- a/docs/zh-CN/spec/03-runtime/04-data-storage.md
+++ b/docs/zh-CN/spec/03-runtime/04-data-storage.md
@@ -1,4 +1,4 @@
-# 04. 数据存储（架构 v23）
+# 04. 数据存储（架构 v25）
 
 > **翻译说明：** 本页是与 [英文源规格](/spec/03-runtime/04-data-storage) 一一对应的机器辅助翻译。代码、协议字段和标识符保持原文；如翻译与英文源事实有歧义，以英文版本为准。
 
@@ -362,6 +362,8 @@ CREATE TABLE sessions (
                                           'high', 'xhigh', 'max', 'omit')),
   permission_mode TEXT NOT NULL DEFAULT 'inherit' -- D115: inherit follows settings
                 CHECK (permission_mode IN ('inherit', 'ask', 'accept-edits', 'auto')),
+  fast_mode   INTEGER NOT NULL DEFAULT 0       -- ADR 0332: session /fast switch
+               CHECK (fast_mode IN (0, 1)),
   source      TEXT,                            -- import origin: claude-code | codex | opencode | pi
   deleted_at  INTEGER,                         -- plugin trash marker; null means active
   pinned      INTEGER NOT NULL DEFAULT 0,
@@ -1160,7 +1162,7 @@ outbox 排空。渲染器侧的停止绝不重写已有已开始回复的转录
 - JSON 列在热路径上盲读（按原样发送到渲染器）；
   任何过滤或求和的内容都是按规则提升的列。
 
-## 7. 版本控制、v7 重置和 v8 到 v23 迁移
+## 7. 版本控制、v7 重置和 v8 到 v25 迁移
 
 - `PRAGMA user_version` 保留模式权限；未来的结构性变化
   再次添加有序的 Rust 迁移 fns，每个都在一个事务中，并带有一个
@@ -1171,7 +1173,7 @@ outbox 排空。渲染器侧的停止绝不重写已有已开始回复的转录
   旧文件中的会话、提供程序和设置不会保留；
   存档仍保留以供手动恢复。所有 v7 之前的迁移代码
   （v1 `settings.sqlite` 导入，v2→v6 链）被删除。
-- 全新安装直接运行完整的 v23 DDL。
+- 全新安装直接运行完整的 v25 DDL。
 - **架构 v15 是增量的。** 它增加 `turn_queue` 表及其两个索引（D386 / ADR 0213），使 Host
   拥有的回合队列在重启后存活；不改动任何已有行，迁移前保留 `pi.sqlite.v14.bak`。
 - **架构 v16 是增量的。** 它增加会话协作 link 和投递表、生命周期索引，以及可为空的
@@ -1184,6 +1186,11 @@ outbox 排空。渲染器侧的停止绝不重写已有已开始回复的转录
   `default`；其余现有标题保持原样并归类为 `manual`。迁移前保留 v22 备份。
   v15→v16 会话协作步骤现在写入 `16`（它自己的版本）而不是最新的架构常量，
   因此 v15 文件可以在一次启动中走完两个步骤。
+- **架构 v24 是增量的。** 它为隔离会话增加 `worktrees` 登记表和
+  `projects.worktree_setup_command`（ADR 0329）。迁移前保留 v23 备份。
+- **架构 v25 是增量的。** 它增加 `sessions.fast_mode`（ADR 0332），即会话的 `/fast`
+  开关，所有已有会话默认为 `0`。fork 会复制该值；定时任务运行和委托子会话从 `0` 开始。
+  迁移前保留 v24 备份。
 - **架构 v7 首先到达 v8，然后使用受保护的路径。** v7→v8
   迁移之后是相同的受保护的 v8→v15 迁移；架构-v9 和
   schema-v10 数据库采用相同的受保护路径并接收精确的可读数据
diff --git a/docs/zh-CN/spec/03-runtime/06-host-rpc-protocol.md b/docs/zh-CN/spec/03-runtime/06-host-rpc-protocol.md
index b471398468..fa7897617b 100644
--- a/docs/zh-CN/spec/03-runtime/06-host-rpc-protocol.md
+++ b/docs/zh-CN/spec/03-runtime/06-host-rpc-protocol.md
@@ -226,7 +226,8 @@ type ToolBudgetHealth = {
 ### 会议
 - `session.list`
 - `session.create` — 接受可选的 `thinkingLevel`； missing/null 默认值
-至 `off`。可选的 `isolated: true`（ADR 0329）同时要求 `projectPath`，并在返回前为新会话
+至 `off`。可选布尔值 `fastMode`（ADR 0332）保存会话的 `/fast` 开关；缺省/null 表示
+`false`，非布尔值返回 `INVALID_PARAMS`。可选的 `isolated: true`（ADR 0329）同时要求 `projectPath`，并在返回前为新会话
 创建专属 git worktree；无法创建时会话会被删除，调用以 `WORKTREE_*` 错误（§7）失败。
 - `session.fork` — accepts `sessionId`, an optional caller-provided display
   `title`, and optional `throughMessageId`; creates
@@ -252,7 +253,8 @@ type ToolBudgetHealth = {
   并且不更新 `updated_at`、转录内容或消息数。
 - `session.configure` — 以原子方式持久保存 `mode`、`providerId`、`modelId`，
   以及可选的 `thinkingLevel` 用于下一个 pi 回合； omitting/null
-  `thinkingLevel` 保留当前值；返回无效模式或级别
+  `thinkingLevel` 保留当前值；可选布尔值 `fastMode`（ADR 0332）遵循同一规则，
+  非布尔值返回 `INVALID_PARAMS`；返回无效模式或级别
   `INVALID_PARAMS`；模式为 `plan | goal | agent` 并更改任何会话
   仅在空闲且没有 pending/queued/running 时才允许配置
   Plan 或 Goal 记录
diff --git a/docs/zh-CN/spec/03-runtime/11-provider-model-system.md b/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
index dec4160c99..b97b732111 100644
--- a/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
+++ b/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
@@ -295,6 +295,7 @@ type ModelBinding = {
   thinkingLevels: ThinkingLevel[]
   defaultThinkingLevel: SessionThinkingLevel | null
   availableForSubagents?: boolean // opt-in for AI-driven delegation
+  fastModeAllowed?: boolean // opt-in for the session /fast switch (ADR 0332)
 }
 
 type SelectedModelRef = {
@@ -323,6 +324,14 @@ type ThinkingLevel =
 后续工作。只有 `defaultModelId` 的旧版提供商，在主机读取时会被具体化为一条
 回退 binding，并在下一次提供商写入时升级为 `models`。
 
+`ModelBinding.fastModeAllowed`（布尔值，默认 false，ADR 0332）：允许会话的 `/fast`
+开关为该模型请求 OpenAI 优先服务档位（`service_tier: "priority"`）。模型设置只在
+Responses 协议上启用该勾选框：`openai_codex_responses`（ChatGPT 订阅，优先档位会更快
+消耗订阅额度）和 `responses`（按优先档位更高的 token 单价计费；中转服务可能忽略或
+另行计费）。只有当会话开关打开且保存的 binding 允许时，启动器才向 sidecar 发送
+`fastMode: true`；运行时只在解析后的协议为 `openai-codex-responses` 或
+`openai-responses` 的主回合请求中加入该字段。子代理、压缩和一次性请求保持默认档位。
+
 `ModelBinding.availableForSubagents`（布尔值，默认 false）：这是一个选择加入
 的标志，让该模型可用于 AI 驱动的子代理委托。启用后，该模型会出现在注入父
 agent 系统提示的委托目录中。父 agent 随后就能通过 Task 工具的 `model` 参数
diff --git a/docs/zh-CN/spec/04-ux/04-builtin-commands.md b/docs/zh-CN/spec/04-ux/04-builtin-commands.md
index 331ee02c80..1403d9266f 100644
--- a/docs/zh-CN/spec/04-ux/04-builtin-commands.md
+++ b/docs/zh-CN/spec/04-ux/04-builtin-commands.md
@@ -4,7 +4,7 @@
 
 ## 1. 目标
 
-定义无需插件即可使用的五条第一方命令。它们也会出现在输入框的 `/` 菜单中。
+定义无需插件即可使用的六条第一方命令。它们也会出现在输入框的 `/` 菜单中。
 
 快捷键：**Cmd/Ctrl + Shift + P**（D014）
 
@@ -16,7 +16,7 @@ builtin.<domain>.<action>
 
 ## 3. 核心内置目录
 
-内置 registry 有意只包含以下五条命令。插件命令会动态扩展可搜索的命令列表；
+内置 registry 有意只包含以下六条命令。插件命令会动态扩展可搜索的命令列表；
 应用导航、项目管理、设置、插件管理和诊断不属于内置命令契约。
 
 | id | 标题 | 关键词 | 类别 | 风险 | 行为 |
@@ -26,10 +26,11 @@ builtin.<domain>.<action>
 | `builtin.mode.agent` | 切换到 Agent | mode、agent | Session | 低 | 将空闲会话模式设为 Agent |
 | `builtin.mode.plan` | 切换到 Plan | mode、plan、planning | Session | 低 | 将空闲会话模式设为 Plan |
 | `builtin.mode.goal` | 切换到 Goal | mode、goal、objective、autonomous | Session | 低 | 将空闲会话模式设为 Goal |
+| `builtin.session.fast` | 切换 Fast 模式 | fast、speed、priority、service tier | Session | 中 | 切换会话的 OpenAI 优先服务档位（ADR 0332） |
 
 ## 4. 可见性与执行规则
 
-- 五个 ID 构成完整的第一方 registry。移除的 ID 不会出现在命令面板结果中，
+- 六个 ID 构成完整的第一方 registry。移除的 ID 不会出现在命令面板结果中，
   也不再有渲染器 dispatch case；插件命令仍可独立发现。
 - `New Task` 使用当前项目或临时分组。若该组最新会话为空则选中复用；否则
   在第一帧露出空首页并创建持久空会话。同一分组内该操作是幂等的。
@@ -38,6 +39,12 @@ builtin.<domain>.<action>
 - 模式命令使用与 Composer Agent/Plan/Goal 芯片相同的活动会话配置路径，立即更新
   空闲会话。没有活动会话时，它们更新下一个会话使用的持久化默认值；正在运行的会话
   或待审批会话不会被修改。
+- `Toggle Fast mode` 通过同一条配置路径翻转会话持久化的 `fastMode`（ADR 0332）。
+  没有活动会话时，它记录在新任务草稿上，创建出的会话会带上该开关。所选模型必须在
+  「设置 → 模型 → 高级」中允许 Fast 模式，且走 Responses 协议（Codex 订阅或 OpenAI
+  Responses）；否则无论开启还是关闭，都只弹出警告指向该设置，不保存任何状态。未配置
+  模型上残留的开关在启动时不生效。远程会话拒绝该命令。风险为中，
+  因为优先档位会更快消耗订阅额度，或按更高的 token 单价计费。
 - `SubmitPlan` 和 `SubmitGoal` 是模型工具，不是命令面板命令。没有 Chat 模式或
   request-changes 别名。
 - 原有的应用、项目、设置、插件和日志操作（仍适用时）通过各自的专用界面提供，
@@ -55,7 +62,7 @@ type CommandExecutionResult =
 
 ## 6. 验收
 
-1. 内置 registry 恰好包含五个唯一且带前缀的 ID。
+1. 内置 registry 恰好包含六个唯一且带前缀的 ID。
 2. 命令面板搜索可匹配每条命令的标题和关键词。
 3. 模式切换命令立即更新空闲会话模式；Plan、Goal 和 Agent 使用同一个 pi Agent。
 4. Compact 在空闲时可用，在活动回合/检查点期间返回 `AGENT_BUSY`。
@@ -74,9 +81,10 @@ type CommandExecutionResult =
 | `/agent-mode` | `builtin.mode.agent` |
 | `/plan-mode` | `builtin.mode.plan` |
 | `/goal-mode` | `builtin.mode.goal` |
+| `/fast` | `builtin.session.fast` |
 
 别名与模板和插件命令名称共享一个命名空间；冲突时优先使用内置别名，其次是项目模板、
-用户模板和插件命令。选择别名会插入 `/alias `；单独发送 `/new` 或 `/compact` 时，
+用户模板和插件命令。选择别名会插入 `/alias `；单独发送 `/new`、`/compact` 或 `/fast` 时，
 会在本地执行，不会创建空提示。Agent/Plan/Goal 别名支持附带提示词：
 `/agent-mode <prompt>`、`/plan-mode <prompt>` 或 `/goal-mode <prompt>` 会切换空闲会话
 （或下一个会话默认值），并通过正常提示路径发送 `<prompt>`。不带正文的模式别名仍然
diff --git a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
index 6783212d8f..492b109528 100644
--- a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
@@ -9765,3 +9765,21 @@ host-created files. The full app's file-preview viewer is covered separately.
 - **关联规格**：`04-ux/01-ui-ia.md` §3.2、`04-ux/06-settings-ia.md`（常规 → 侧边栏卡片）。
 - **验收**：项目折叠时仍能回答“这里是否有正在运行的工作”，且两种提示可从持久化设置独立控制。
 - **状态**：草稿。现有 `apps/desktop/test/sidebar-project-runs.test.mjs` 覆盖数量分类、总开关与子开关门控、两种渲染入口，以及主题令牌和减少动态效果保证；它不覆盖设置 API 缺省值规范化或完整交互，相关场景仍需在可用环境手动验证。
+
+### E2E-FAST-MODE：`/fast` 只在允许的模型上请求优先档位
+
+- **前置条件**：一个 ChatGPT 订阅（Codex）账号和一个 Responses API key 提供商，各配置两个模型。
+  自动化路径不需要真实提供商；手动路径使用可捕获请求的本地端点或用户授权的账号。
+- **步骤**：
+  1. 在「设置 → 模型」中为每个提供商的一个模型勾选 **Fast 模式**；确认 Chat Completions 提供商上该勾选框不可用。
+  2. 在允许的 Codex 模型上新建任务，输入 `/fast`，发送一条提示。
+  3. 再次发送 `/fast`，然后再发送一条提示。
+  4. 把会话切换到未允许的模型，发送 `/fast`。
+  5. 在 Fast 开启时 fork 会话；重启应用后重新打开两个会话。
+- **预期**：第 1 步只在 binding 上保存 `fastModeAllowed`。第 2 步模型芯片显示 Fast 标识，请求体带有
+  `service_tier: "priority"`；Responses API key 模型改为提示按 token 计费的价格。第 3 步清除标识和该字段。
+  第 4 步给出模型设置提示并拒绝，不保存任何状态。第 5 步 fork 和重启后开关保持开启。子代理和压缩请求
+  从不携带该字段。远程会话拒绝 `/fast`。
+- **关联规格**：`03-runtime/11-provider-model-system.md`、`03-runtime/04-data-storage.md`（v25）、
+  `04-ux/04-builtin-commands.md`、ADR 0332、issue #1657。
+- **状态**：自动化覆盖见英文版。完整的设置 → `/fast` → 捕获请求路径为手动验证场景。
diff --git a/packages/agent-runtime/src/fast-mode.test.ts b/packages/agent-runtime/src/fast-mode.test.ts
new file mode 100644
index 0000000000..0bf505f499
--- /dev/null
+++ b/packages/agent-runtime/src/fast-mode.test.ts
@@ -0,0 +1,75 @@
+import { describe, expect, it } from "vitest";
+import type { Model, SimpleStreamOptions } from "@earendil-works/pi-ai";
+import { streamSimple as completionsStreamSimple } from "@earendil-works/pi-ai/api/openai-completions";
+import { streamSimple as responsesStreamSimple } from "@earendil-works/pi-ai/api/openai-responses";
+import { withFastModeOptions } from "./fast-mode.js";
+
+const RELAY = "https://relay.example.com/v1";
+
+const responsesModel = {
+  id: "gpt-test",
+  name: "GPT Test",
+  api: "openai-responses",
+  provider: "relay-uuid",
+  baseUrl: RELAY,
+  reasoning: false,
+  input: ["text"],
+  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+  contextWindow: 128_000,
+  maxTokens: 8_192,
+} as Model<"openai-responses">;
+
+function capture() {
+  const bodies: Array<Record<string, unknown>> = [];
+  const fetch: typeof globalThis.fetch = async (_url, init) => {
+    bodies.push(JSON.parse(String(init?.body)) as Record<string, unknown>);
+    return new Response(JSON.stringify({ error: { message: "fixture response" } }), {
+      status: 502,
+      headers: { "content-type": "application/json" },
+    });
+  };
+  return { bodies, options: { apiKey: "sk-test", fetch } satisfies SimpleStreamOptions };
+}
+
+const context = { messages: [{ role: "user", content: "hi", timestamp: 1 }] } as never;
+
+describe("withFastModeOptions", () => {
+  it("sends the priority tier on the Responses wire when allowed and switched on", async () => {
+    const { bodies, options } = capture();
+    await responsesStreamSimple(
+      responsesModel,
+      context,
+      withFastModeOptions(options, true),
+    ).result();
+    expect(bodies).toHaveLength(1);
+    expect(bodies[0].service_tier).toBe("priority");
+  });
+
+  it("keeps the default tier when Fast mode is off", async () => {
+    const { bodies, options } = capture();
+    const hooked = withFastModeOptions(options, false);
+    expect(hooked).toBe(options);
+    await responsesStreamSimple(responsesModel, context, hooked).result();
+    expect(bodies[0].service_tier).toBeUndefined();
+  });
+
+  it("leaves a Chat Completions request untouched even when switched on", async () => {
+    const { bodies, options } = capture();
+    const model = { ...responsesModel, api: "openai-completions" } as Model<"openai-completions">;
+    await completionsStreamSimple(
+      model,
+      context,
+      withFastModeOptions(options, true),
+    ).result();
+    expect(bodies[0].service_tier).toBeUndefined();
+  });
+
+  it("chains after an earlier payload hook and keeps its replacement", async () => {
+    const options: SimpleStreamOptions = {
+      onPayload: async (payload) => ({ ...(payload as object), prompt_cache_key: "kept" }),
+    };
+    const hooked = withFastModeOptions(options, true);
+    const payload = await hooked.onPayload?.({ model: "gpt-test" }, responsesModel);
+    expect(payload).toEqual({ model: "gpt-test", prompt_cache_key: "kept", service_tier: "priority" });
+  });
+});
diff --git a/packages/agent-runtime/src/fast-mode.ts b/packages/agent-runtime/src/fast-mode.ts
new file mode 100644
index 0000000000..893bc58a16
--- /dev/null
+++ b/packages/agent-runtime/src/fast-mode.ts
@@ -0,0 +1,34 @@
+/**
+ * OpenAI Fast mode on the request body (ADR 0332).
+ *
+ * pi-ai's `streamSimple` does not forward a service tier, so the priority tier
+ * is set through the payload hook, the same channel the session-affinity body
+ * key uses. The launcher has already combined the session `/fast` switch with
+ * the model binding's opt-in; the hook still reads the wire of the model being
+ * called, so a request that does not travel on a Responses wire keeps its bytes.
+ */
+
+import type { SimpleStreamOptions } from "@earendil-works/pi-ai";
+import { FAST_MODE_SERVICE_TIER, fastModeSupportedOn } from "@pi-desktop/shared";
+
+/** Chain a payload hook that asks for the priority tier when Fast mode is on. */
+export function withFastModeOptions(
+  options: SimpleStreamOptions,
+  fastMode: boolean,
+): SimpleStreamOptions {
+  if (fastMode !== true) return options;
+  const previous = options.onPayload;
+  return {
+    ...options,
+    onPayload: async (payload, model) => {
+      const replacement = await previous?.(payload, model);
+      const base = replacement === undefined ? payload : replacement;
+      if (!isRecord(base) || !fastModeSupportedOn(model.api)) return replacement;
+      return { ...base, service_tier: FAST_MODE_SERVICE_TIER };
+    },
+  };
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
diff --git a/packages/agent-runtime/src/runtime.ts b/packages/agent-runtime/src/runtime.ts
index aeeeedb62a..5c576855a1 100644
--- a/packages/agent-runtime/src/runtime.ts
+++ b/packages/agent-runtime/src/runtime.ts
@@ -240,6 +240,7 @@ import {
   openCodeEndpointFromProvider,
   withOpenCodeSessionHeaders,
 } from "./opencode-session-headers.js";
+import { withFastModeOptions } from "./fast-mode.js";
 import { withSessionAffinityOptions } from "./session-affinity.js";
 import { withCompactionRequestHeaders } from "./compaction-request.js";
 import type { CompactionRequestShape } from "./compaction-diagnostics.js";
@@ -978,6 +979,11 @@ export type AgentRuntimeOptions = {
   thinkingLevel: SessionThinkingLevel;
   /** Persisted opt-in for retrying transient provider failures until success. */
   infiniteProviderRetry?: boolean;
+  /**
+   * Fast mode for main turns, already resolved by the launcher from the
+   * session `/fast` switch and the model binding's opt-in (ADR 0332).
+   */
+  fastMode?: boolean;
   /** TypeSafe Jev credential, resolved by Electron main only when enabled. */
   jevApiKey?: string;
   systemPrompt?: string;
@@ -1920,6 +1926,7 @@ export class DesktopAgentRuntime {
   private providerRateLimitRetryAttempt = 0;
   /** Opt-in mode removes only the retry-count ceiling; abort and backoff stay intact. */
   private infiniteProviderRetry = false;
+  private fastMode = false;
   private jevApiKey?: string;
   private activeProviderRetryAttempt = 0;
   private providerRetryInProgress = false;
@@ -2023,6 +2030,7 @@ export class DesktopAgentRuntime {
     this.provider = opts.provider;
     this.thinkingLevel = clampThinkingLevel(opts.provider, opts.thinkingLevel);
     this.infiniteProviderRetry = opts.infiniteProviderRetry === true;
+    this.fastMode = opts.fastMode === true;
     this.jevApiKey = opts.jevApiKey;
     this.host = opts.host;
     this.onDiagnostic = opts.onDiagnostic ?? (() => undefined);
@@ -2183,14 +2191,19 @@ Do not invent objections or turn speculative risks into blockers. Stop when the
           ),
           m.api,
         );
-        const hookedOptions = withSessionAffinityOptions(
-          this.withExtensionProviderHooks(requestOptions, m),
-          {
-            provider: this.provider,
-            api: m.api,
-            baseUrl: m.baseUrl,
-            sessionId: this.sessionId,
-          },
+        // Only main turns ask for the priority tier: subagents, compaction and
+        // one-shot requests build their own options and keep the default.
+        const hookedOptions = withFastModeOptions(
+          withSessionAffinityOptions(
+            this.withExtensionProviderHooks(requestOptions, m),
+            {
+              provider: this.provider,
+              api: m.api,
+              baseUrl: m.baseUrl,
+              sessionId: this.sessionId,
+            },
+          ),
+          this.fastMode,
         );
         // The watchdog must be able to *stop* what it abandons. It wraps the
         // retry adapter, so draining alone would let the adapter wake from its
@@ -2329,6 +2342,12 @@ Do not invent objections or turn speculative risks into blockers. Stop when the
     this.infiniteProviderRetry = enabled;
   }
 
+  /** Apply the session `/fast` switch to the next turn of a reused runtime. */
+  setFastMode(enabled: boolean): void {
+    if (this.disposed) throw new Error("runtime disposed");
+    this.fastMode = enabled;
+  }
+
   /** Refresh catalog instructions without changing the running session owner. */
   setPluginSkills(skills: PluginSkillDef[]): void {
     if (this.disposed) throw new Error("runtime disposed");
diff --git a/packages/agent-runtime/src/sidecar.ts b/packages/agent-runtime/src/sidecar.ts
index c591dabaf4..9ed694ee00 100644
--- a/packages/agent-runtime/src/sidecar.ts
+++ b/packages/agent-runtime/src/sidecar.ts
@@ -106,6 +106,8 @@ type RuntimeParams = {
   turnId?: string;
   thinkingLevel?: SessionThinkingLevel;
   infiniteProviderRetry?: boolean;
+  /** Session `/fast` switch AND the model's opt-in, resolved by main (ADR 0332). */
+  fastMode?: boolean;
   /** Opt-in TypeSafe classifier credential resolved by Electron main. */
   jevApiKey?: string;
   provider: RuntimeProviderConfig;
@@ -253,6 +255,7 @@ async function runtimeFor(
     reusable.setPluginSkills(pluginSkills);
     reusable.setCompactionSettings(params.compactionSettings);
     reusable.setInfiniteProviderRetry(params.infiniteProviderRetry === true);
+    reusable.setFastMode(params.fastMode === true);
     reusable.setMode(mode);
     return reusable;
   }
@@ -301,6 +304,7 @@ async function runtimeFor(
     commandShell: params.commandShell,
     thinkingLevel,
     infiniteProviderRetry: params.infiniteProviderRetry === true,
+    fastMode: params.fastMode === true,
     jevApiKey: params.jevApiKey,
     history,
     compaction,
diff --git a/packages/i18n/src/locales/de/index.ts b/packages/i18n/src/locales/de/index.ts
index 0fa00141cb..5be755158b 100644
--- a/packages/i18n/src/locales/de/index.ts
+++ b/packages/i18n/src/locales/de/index.ts
@@ -645,6 +645,13 @@ export const de = {
     "modelBadgeVision": "Vision",
     "permissionMode": "Berechtigungsmodus",
     "reasoningLevel": "Argumentationsebene",
+    "fastModeBadge": "Fast",
+    "fastModeOn": "Fast-Modus für diese Sitzung an. Antworten kommen schneller und verbrauchen mehr Nutzungslimit deines Abos.",
+    "fastModeOnUsage": "Fast-Modus für diese Sitzung an. Dieses Modell wird pro Token zum höheren Preis der Prioritätsstufe abgerechnet.",
+    "fastModeOff": "Fast-Modus für diese Sitzung aus.",
+    "fastModeUnsupported": "Der Fast-Modus braucht ein Modell über die OpenAI Responses API oder ein ChatGPT-Abo.",
+    "fastModeNotAllowed": "Für dieses Modell ist der Fast-Modus nicht eingerichtet. Aktiviere ihn unter Einstellungen → Modelle → Erweitert.",
+    "fastModeRemote": "Remote-Sitzungen nutzen die Modellkonfiguration des Hosts, daher ist der Fast-Modus nicht verfügbar.",
     "permissionInherit": "Standard",
     "permissionAsk": "Immer nachfragen",
     "permissionAcceptEdits": "Änderungen akzeptieren",
@@ -1712,6 +1719,10 @@ sklm: {
     "nativeWebSearch": "Native Websuche",
     "nativeWebSearchHint": "Hängt das vom Anbieter gehostete Websuche-Tool an Anfragen für dieses Modell an. Erfordert einen Endpunkt, der dies unterstützt; die Abrechnung erfolgt pro Suche beim Anbieter.",
     nativeWebSearchUnsupported: "Diese Verbindung unterstützt in der App noch keine native Suche. Die Website oder eine andere API kann Suche anbieten.",
+    fastModeAllowed: "Fast-Modus",
+    fastModeAllowedSubscriptionHint: "Erlaubt /fast, für dieses Modell die Prioritätsstufe von OpenAI anzufordern. Antworten kommen schneller und verbrauchen mehr Nutzungslimit deines ChatGPT-Abos.",
+    fastModeAllowedUsageHint: "Erlaubt /fast, für dieses Modell die Prioritätsstufe von OpenAI (service_tier) anzufordern. OpenAI berechnet dafür einen höheren Preis pro Token; ein Relay kann das Feld ignorieren oder anders abrechnen.",
+    fastModeAllowedUnsupported: "Der Fast-Modus braucht die OpenAI Responses API oder ein ChatGPT-Abo.",
     "searchModelId": "Modell-ID suchen…",
     "searchChosenModels": "Hinzugefügte Modelle suchen…",
     "noChosenModelMatches": "Keine passenden hinzugefügten Modelle.",
diff --git a/packages/i18n/src/locales/en/index.ts b/packages/i18n/src/locales/en/index.ts
index 7b16fca985..f6987ac412 100644
--- a/packages/i18n/src/locales/en/index.ts
+++ b/packages/i18n/src/locales/en/index.ts
@@ -657,6 +657,13 @@ export const en = {
     modelBadgeVision: "vision",
     permissionMode: "Permission mode",
     reasoningLevel: "Reasoning level",
+    fastModeBadge: "Fast",
+    fastModeOn: "Fast mode on for this session. Replies come faster and use more of your plan's usage limits.",
+    fastModeOnUsage: "Fast mode on for this session. This model is billed per token at the priority tier's higher price.",
+    fastModeOff: "Fast mode off for this session.",
+    fastModeUnsupported: "Fast mode needs a model on the OpenAI Responses API or a ChatGPT subscription.",
+    fastModeNotAllowed: "Fast mode is not configured for this model. Turn it on in Settings → Models → Advanced.",
+    fastModeRemote: "Remote sessions use the Host's model configuration, so Fast mode is not available.",
     permissionInherit: "Default",
     permissionAsk: "Ask every time",
     permissionAcceptEdits: "Accept edits",
@@ -1732,6 +1739,10 @@ sklm: {
     nativeWebSearch: "Native web search",
     nativeWebSearchHint: "Attach the provider-hosted web search tool to requests for this model. Requires an endpoint that supports it; billed by the provider per search.",
     nativeWebSearchUnsupported: "This app has not integrated native search for this connection. Website search or another API may still be available.",
+    fastModeAllowed: "Fast mode",
+    fastModeAllowedSubscriptionHint: "Let /fast request OpenAI's priority tier for this model. Replies come faster and use more of your ChatGPT plan's usage limits.",
+    fastModeAllowedUsageHint: "Let /fast request OpenAI's priority tier (service_tier) for this model. OpenAI bills the tier at a higher per-token price; a relay may ignore the field or bill it differently.",
+    fastModeAllowedUnsupported: "Fast mode needs the OpenAI Responses API or a ChatGPT subscription.",
     searchModelId: "Search model ID…",
     searchChosenModels: "Search added models…",
     noChosenModelMatches: "No matching added models.",
diff --git a/packages/i18n/src/locales/es/index.ts b/packages/i18n/src/locales/es/index.ts
index 205a5595ad..353b5ca8a0 100644
--- a/packages/i18n/src/locales/es/index.ts
+++ b/packages/i18n/src/locales/es/index.ts
@@ -645,6 +645,13 @@ export const es = {
     "modelBadgeVision": "visión",
     "permissionMode": "Modo de permiso",
     "reasoningLevel": "Nivel de razonamiento",
+    "fastModeBadge": "Fast",
+    "fastModeOn": "Modo Fast activado en esta sesión. Las respuestas llegan antes y consumen más del límite de uso de tu plan.",
+    "fastModeOnUsage": "Modo Fast activado en esta sesión. Este modelo se factura por token al precio más alto del nivel prioritario.",
+    "fastModeOff": "Modo Fast desactivado en esta sesión.",
+    "fastModeUnsupported": "El modo Fast necesita un modelo de la API Responses de OpenAI o una suscripción a ChatGPT.",
+    "fastModeNotAllowed": "El modo Fast no está configurado para este modelo. Actívalo en Ajustes → Modelos → Avanzado.",
+    "fastModeRemote": "Las sesiones remotas usan la configuración de modelos del host, así que el modo Fast no está disponible.",
     "permissionInherit": "Predeterminado",
     "permissionAsk": "Preguntar siempre",
     "permissionAcceptEdits": "Aceptar ediciones",
@@ -1712,6 +1719,10 @@ sklm: {
     "nativeWebSearch": "Búsqueda web nativa",
     "nativeWebSearchHint": "Adjunta la herramienta de búsqueda web alojada por el proveedor a las solicitudes de este modelo. Requiere un punto de conexión que la admita; el proveedor la factura por búsqueda.",
     nativeWebSearchUnsupported: "La aplicación aún no integra búsqueda nativa para esta conexión. El sitio web u otra API pueden ofrecerla.",
+    fastModeAllowed: "Modo Fast",
+    fastModeAllowedSubscriptionHint: "Permite que /fast pida el nivel prioritario de OpenAI para este modelo. Las respuestas llegan antes y consumen más del límite de uso de tu plan de ChatGPT.",
+    fastModeAllowedUsageHint: "Permite que /fast pida el nivel prioritario de OpenAI (service_tier) para este modelo. OpenAI lo factura a un precio por token más alto; un relay puede ignorar el campo o facturarlo de otra forma.",
+    fastModeAllowedUnsupported: "El modo Fast necesita la API Responses de OpenAI o una suscripción a ChatGPT.",
     "searchModelId": "ID de modelo de búsqueda…",
     "searchChosenModels": "Buscar modelos añadidos…",
     "noChosenModelMatches": "No hay modelos añadidos que coincidan.",
diff --git a/packages/i18n/src/locales/fr/index.ts b/packages/i18n/src/locales/fr/index.ts
index f015a56042..d591515ff7 100644
--- a/packages/i18n/src/locales/fr/index.ts
+++ b/packages/i18n/src/locales/fr/index.ts
@@ -645,6 +645,13 @@ export const fr = {
     "modelBadgeVision": "vision",
     "permissionMode": "Mode d'autorisation",
     "reasoningLevel": "Niveau de raisonnement",
+    "fastModeBadge": "Fast",
+    "fastModeOn": "Mode Fast activé pour cette session. Les réponses arrivent plus vite et consomment davantage les limites d'utilisation de votre abonnement.",
+    "fastModeOnUsage": "Mode Fast activé pour cette session. Ce modèle est facturé au token, au tarif plus élevé du niveau prioritaire.",
+    "fastModeOff": "Mode Fast désactivé pour cette session.",
+    "fastModeUnsupported": "Le mode Fast nécessite un modèle sur l'API Responses d'OpenAI ou un abonnement ChatGPT.",
+    "fastModeNotAllowed": "Le mode Fast n'est pas configuré pour ce modèle. Activez-le dans Réglages → Modèles → Avancé.",
+    "fastModeRemote": "Les sessions distantes utilisent la configuration des modèles de l'hôte : le mode Fast n'est pas disponible.",
     "permissionInherit": "Par défaut",
     "permissionAsk": "Demander à chaque fois",
     "permissionAcceptEdits": "Accepter les modifications",
@@ -1712,6 +1719,10 @@ sklm: {
     "nativeWebSearch": "Recherche web native",
     "nativeWebSearchHint": "Attache l'outil de recherche web hébergé par le fournisseur aux requêtes pour ce modèle. Nécessite un point de terminaison qui le prend en charge ; facturé par le fournisseur à chaque recherche.",
     nativeWebSearchUnsupported: "La recherche native n’est pas encore intégrée pour cette connexion. Le site web ou une autre API peut la proposer.",
+    fastModeAllowed: "Mode Fast",
+    fastModeAllowedSubscriptionHint: "Autorise /fast à demander le niveau prioritaire d'OpenAI pour ce modèle. Les réponses arrivent plus vite et consomment davantage les limites de votre abonnement ChatGPT.",
+    fastModeAllowedUsageHint: "Autorise /fast à demander le niveau prioritaire d'OpenAI (service_tier) pour ce modèle. OpenAI le facture à un prix par token plus élevé ; un relais peut ignorer ce champ ou le facturer autrement.",
+    fastModeAllowedUnsupported: "Le mode Fast nécessite l'API Responses d'OpenAI ou un abonnement ChatGPT.",
     "searchModelId": "ID de modèle de recherche…",
     "searchChosenModels": "Rechercher les modèles ajoutés…",
     "noChosenModelMatches": "Aucun modèle ajouté correspondant.",
diff --git a/packages/i18n/src/locales/ko/index.ts b/packages/i18n/src/locales/ko/index.ts
index 5c5b1723f9..b22559848d 100644
--- a/packages/i18n/src/locales/ko/index.ts
+++ b/packages/i18n/src/locales/ko/index.ts
@@ -654,6 +654,13 @@ export const ko = {
     modelBadgeVision: "비전",
     permissionMode: "권한 모드",
     reasoningLevel: "추론 수준",
+    fastModeBadge: "Fast",
+    fastModeOn: "이 세션에서 Fast 모드를 켰습니다. 응답이 더 빨라지고 구독 사용 한도를 더 많이 소모합니다.",
+    fastModeOnUsage: "이 세션에서 Fast 모드를 켰습니다. 이 모델은 우선 등급의 더 높은 토큰 단가로 청구됩니다.",
+    fastModeOff: "이 세션에서 Fast 모드를 껐습니다.",
+    fastModeUnsupported: "Fast 모드는 OpenAI Responses API 또는 ChatGPT 구독 모델에서만 사용할 수 있습니다.",
+    fastModeNotAllowed: "이 모델에는 Fast 모드가 설정되어 있지 않습니다. 설정 → 모델 → 고급에서 켜세요.",
+    fastModeRemote: "원격 세션은 호스트의 모델 구성을 사용하므로 Fast 모드를 사용할 수 없습니다.",
     permissionInherit: "기본값",
     permissionAsk: "매번 묻기",
     permissionAcceptEdits: "편집 허용",
@@ -1728,6 +1735,10 @@ sklm: {
     nativeWebSearch: "네이티브 웹 검색",
     nativeWebSearchHint: "이 모델에 대한 요청에 공급자 호스팅 웹 검색 도구를 첨부합니다. 이를 지원하는 엔드포인트가 필요하며, 검색 건수로 공급자에게 과금됩니다.",
     nativeWebSearchUnsupported: "이 연결의 네이티브 검색은 아직 앱에 통합되지 않았습니다. 웹사이트나 다른 API에서는 검색을 지원할 수 있습니다.",
+    fastModeAllowed: "Fast 모드",
+    fastModeAllowedSubscriptionHint: "/fast가 이 모델에 OpenAI 우선 등급을 요청하도록 허용합니다. 응답이 더 빨라지고 ChatGPT 구독 사용 한도를 더 많이 소모합니다.",
+    fastModeAllowedUsageHint: "/fast가 이 모델에 OpenAI 우선 등급(service_tier)을 요청하도록 허용합니다. OpenAI는 더 높은 토큰 단가로 청구하며, 중계 서비스는 이 필드를 무시하거나 다르게 청구할 수 있습니다.",
+    fastModeAllowedUnsupported: "Fast 모드는 OpenAI Responses API 또는 ChatGPT 구독이 필요합니다.",
     searchModelId: "모델 ID 검색…",
     searchChosenModels: "추가된 모델 검색…",
     noChosenModelMatches: "일치하는 추가된 모델이 없습니다.",
diff --git a/packages/i18n/src/locales/pt-BR/index.ts b/packages/i18n/src/locales/pt-BR/index.ts
index 48036f4236..5c544db87d 100644
--- a/packages/i18n/src/locales/pt-BR/index.ts
+++ b/packages/i18n/src/locales/pt-BR/index.ts
@@ -644,6 +644,13 @@ export const ptBR = {
     modelBadgeVision: "Visão",
     permissionMode: "Modo de permissão",
     reasoningLevel: "Nível de raciocínio",
+    fastModeBadge: "Fast",
+    fastModeOn: "Modo Fast ativado nesta sessão. As respostas chegam mais rápido e consomem mais do limite de uso do seu plano.",
+    fastModeOnUsage: "Modo Fast ativado nesta sessão. Este modelo é cobrado por token no preço mais alto do nível prioritário.",
+    fastModeOff: "Modo Fast desativado nesta sessão.",
+    fastModeUnsupported: "O modo Fast precisa de um modelo na API Responses da OpenAI ou de uma assinatura do ChatGPT.",
+    fastModeNotAllowed: "O modo Fast não está configurado para este modelo. Ative em Configurações → Modelos → Avançado.",
+    fastModeRemote: "Sessões remotas usam a configuração de modelos do host, então o modo Fast não está disponível.",
     permissionInherit: "Herdar",
     permissionAsk: "Perguntar",
     permissionAcceptEdits: "Aceitar edições",
@@ -1682,6 +1689,10 @@ export const ptBR = {
     nativeWebSearch: "Pesquisa nativa na web",
     nativeWebSearchHint: "Anexa às solicitações deste modelo a ferramenta de pesquisa na web do provedor. Exige um endpoint compatível; o provedor cobra por pesquisa.",
     nativeWebSearchUnsupported: "O aplicativo ainda não integra pesquisa nativa para esta conexão. O site ou outra API ainda pode oferecer pesquisa.",
+    fastModeAllowed: "Modo Fast",
+    fastModeAllowedSubscriptionHint: "Permite que /fast peça o nível prioritário da OpenAI para este modelo. As respostas chegam mais rápido e consomem mais do limite de uso do seu plano do ChatGPT.",
+    fastModeAllowedUsageHint: "Permite que /fast peça o nível prioritário da OpenAI (service_tier) para este modelo. A OpenAI cobra um preço por token mais alto; um relay pode ignorar o campo ou cobrar de outra forma.",
+    fastModeAllowedUnsupported: "O modo Fast precisa da API Responses da OpenAI ou de uma assinatura do ChatGPT.",
     searchModelId: "ID de modelo de pesquisa…",
     searchChosenModels: "Pesquisar modelos adicionados…",
     noChosenModelMatches: "Nenhum modelo escolhido corresponde à busca.",
diff --git a/packages/i18n/src/locales/tr/index.ts b/packages/i18n/src/locales/tr/index.ts
index 10d7331cff..8f2cf24ba7 100644
--- a/packages/i18n/src/locales/tr/index.ts
+++ b/packages/i18n/src/locales/tr/index.ts
@@ -654,6 +654,13 @@ export const tr = {
     modelBadgeVision: "görüntü",
     permissionMode: "İzin kipi",
     reasoningLevel: "Akıl yürütme düzeyi",
+    fastModeBadge: "Fast",
+    fastModeOn: "Bu oturumda Fast modu açık. Yanıtlar daha hızlı gelir ve planınızın kullanım sınırından daha fazla harcar.",
+    fastModeOnUsage: "Bu oturumda Fast modu açık. Bu model, öncelikli katmanın daha yüksek token fiyatıyla ücretlendirilir.",
+    fastModeOff: "Bu oturumda Fast modu kapalı.",
+    fastModeUnsupported: "Fast modu, OpenAI Responses API'sindeki bir model veya ChatGPT aboneliği gerektirir.",
+    fastModeNotAllowed: "Bu model için Fast modu yapılandırılmadı. Ayarlar → Modeller → Gelişmiş bölümünden açın.",
+    fastModeRemote: "Uzak oturumlar ana makinenin model yapılandırmasını kullanır, bu yüzden Fast modu kullanılamaz.",
     permissionInherit: "Varsayılan",
     permissionAsk: "Her seferinde sor",
     permissionAcceptEdits: "Düzenlemeleri kabul et",
@@ -1722,6 +1729,10 @@ sklm: {
     nativeWebSearch: "Yerel web araması",
     nativeWebSearchHint: "Bu modele yapılan isteklere sağlayıcının barındırdığı web arama aracını ekler. Bunu destekleyen bir uç nokta gerekir; arama başına sağlayıcı tarafından faturalandırılır.",
     nativeWebSearchUnsupported: "Bu bağlantı için yerel arama henüz uygulamaya entegre edilmedi. Web sitesi veya başka bir API aramayı destekleyebilir.",
+    fastModeAllowed: "Fast modu",
+    fastModeAllowedSubscriptionHint: "/fast komutunun bu model için OpenAI öncelikli katmanını istemesine izin verir. Yanıtlar daha hızlı gelir ve ChatGPT planınızın kullanım sınırından daha fazla harcar.",
+    fastModeAllowedUsageHint: "/fast komutunun bu model için OpenAI öncelikli katmanını (service_tier) istemesine izin verir. OpenAI bunu daha yüksek token fiyatıyla ücretlendirir; bir aktarıcı alanı yok sayabilir veya farklı ücretlendirebilir.",
+    fastModeAllowedUnsupported: "Fast modu, OpenAI Responses API'si veya ChatGPT aboneliği gerektirir.",
     searchModelId: "Model kimliği ara…",
     searchChosenModels: "Eklenen modellerde ara…",
     noChosenModelMatches: "Eşleşen eklenen model yok.",
diff --git a/packages/i18n/src/locales/zh-CN/index.ts b/packages/i18n/src/locales/zh-CN/index.ts
index b193bfda29..1bcbc7820a 100644
--- a/packages/i18n/src/locales/zh-CN/index.ts
+++ b/packages/i18n/src/locales/zh-CN/index.ts
@@ -650,6 +650,13 @@ export const zhCN = {
     modelBadgeVision: "视觉",
     permissionMode: "权限模式",
     reasoningLevel: "推理等级",
+    fastModeBadge: "Fast",
+    fastModeOn: "已为当前会话开启 Fast 模式：回复更快，但会更快消耗订阅额度。",
+    fastModeOnUsage: "已为当前会话开启 Fast 模式：该模型按优先档位的更高单价按 token 计费。",
+    fastModeOff: "已为当前会话关闭 Fast 模式。",
+    fastModeUnsupported: "Fast 模式需要使用 OpenAI Responses 接口或 ChatGPT 订阅的模型。",
+    fastModeNotAllowed: "该模型未配置 Fast 模式，请到「设置 → 模型 → 高级」中开启。",
+    fastModeRemote: "远程会话使用远程主机的模型配置，无法开启 Fast 模式。",
     permissionInherit: "跟随默认",
     permissionAsk: "每次询问",
     permissionAcceptEdits: "允许编辑",
@@ -1707,6 +1714,10 @@ sklm: {
     nativeWebSearch: "原生联网搜索",
     nativeWebSearchHint: "为该模型的请求附加提供商托管的联网搜索工具。需要端点实际支持；按次计费由提供商收取。",
     nativeWebSearchUnsupported: "应用尚未适配此连接的原生搜索；不代表厂商官网或其他 API 不支持搜索。",
+    fastModeAllowed: "Fast 模式",
+    fastModeAllowedSubscriptionHint: "允许 /fast 为该模型请求 OpenAI 优先档位：回复更快，但会更快消耗 ChatGPT 订阅额度。",
+    fastModeAllowedUsageHint: "允许 /fast 为该模型请求 OpenAI 优先档位（service_tier）。OpenAI 按更高的 token 单价计费；中转服务可能忽略该字段或另行计费。",
+    fastModeAllowedUnsupported: "Fast 模式需要 OpenAI Responses 接口或 ChatGPT 订阅。",
     searchModelId: "搜索模型 ID…",
     searchChosenModels: "搜索已添加模型…",
     noChosenModelMatches: "没有匹配的已添加模型",
diff --git a/packages/i18n/src/locales/zh-TW/index.ts b/packages/i18n/src/locales/zh-TW/index.ts
index a954bec613..6c3d4db80c 100644
--- a/packages/i18n/src/locales/zh-TW/index.ts
+++ b/packages/i18n/src/locales/zh-TW/index.ts
@@ -650,6 +650,13 @@ export const zhTW = {
     modelBadgeVision: "視覺",
     permissionMode: "許可權模式",
     reasoningLevel: "推理等級",
+    fastModeBadge: "Fast",
+    fastModeOn: "已為目前工作階段開啟 Fast 模式：回覆更快，但會更快消耗訂閱額度。",
+    fastModeOnUsage: "已為目前工作階段開啟 Fast 模式：此模型依優先層級的較高單價按 token 計費。",
+    fastModeOff: "已為目前工作階段關閉 Fast 模式。",
+    fastModeUnsupported: "Fast 模式需要使用 OpenAI Responses 介面或 ChatGPT 訂閱的模型。",
+    fastModeNotAllowed: "此模型未設定 Fast 模式，請到「設定 → 模型 → 高階」中開啟。",
+    fastModeRemote: "遠端工作階段使用遠端主機的模型設定，無法開啟 Fast 模式。",
     permissionInherit: "跟隨預設",
     permissionAsk: "每次詢問",
     permissionAcceptEdits: "允許編輯",
@@ -1707,6 +1714,10 @@ sklm: {
     nativeWebSearch: "原生網路搜尋",
     nativeWebSearchHint: "為該模型的請求附加提供商託管的網路搜尋工具。需要端點實際支援；按次計費由提供商收取。",
     nativeWebSearchUnsupported: "應用程式尚未整合此連線的原生搜尋；不代表廠商官網或其他 API 不支援搜尋。",
+    fastModeAllowed: "Fast 模式",
+    fastModeAllowedSubscriptionHint: "允許 /fast 為此模型請求 OpenAI 優先層級：回覆更快，但會更快消耗 ChatGPT 訂閱額度。",
+    fastModeAllowedUsageHint: "允許 /fast 為此模型請求 OpenAI 優先層級（service_tier）。OpenAI 以較高的 token 單價計費；中轉服務可能忽略此欄位或另行計費。",
+    fastModeAllowedUnsupported: "Fast 模式需要 OpenAI Responses 介面或 ChatGPT 訂閱。",
     searchModelId: "搜尋模型 ID…",
     searchChosenModels: "搜尋已新增模型…",
     noChosenModelMatches: "沒有符合的已新增模型",
diff --git a/packages/shared/src/fast-mode.test.ts b/packages/shared/src/fast-mode.test.ts
new file mode 100644
index 0000000000..50604ab782
--- /dev/null
+++ b/packages/shared/src/fast-mode.test.ts
@@ -0,0 +1,19 @@
+import { describe, expect, it } from "vitest";
+import { fastModeBillingOn, fastModeSupportedOn } from "./fast-mode.js";
+
+describe("fast mode wire rule", () => {
+  it("prices the Codex backend on plan limits and Responses per token", () => {
+    expect(fastModeBillingOn("openai_codex_responses")).toBe("subscription");
+    expect(fastModeBillingOn("openai-codex-responses")).toBe("subscription");
+    expect(fastModeBillingOn("responses")).toBe("usage");
+    expect(fastModeBillingOn(" Responses ")).toBe("usage");
+    expect(fastModeBillingOn("openai-responses")).toBe("usage");
+  });
+
+  it("rejects wires that do not carry service_tier", () => {
+    for (const api of ["chat_completions", "openai-completions", "anthropic_messages",
+      "azure-openai-responses", "opencode_go", "", undefined, null]) {
+      expect(fastModeSupportedOn(api)).toBe(false);
+    }
+  });
+});
diff --git a/packages/shared/src/fast-mode.ts b/packages/shared/src/fast-mode.ts
new file mode 100644
index 0000000000..d6d4078ac0
--- /dev/null
+++ b/packages/shared/src/fast-mode.ts
@@ -0,0 +1,26 @@
+/**
+ * OpenAI Fast mode (ADR 0332): the Responses body field
+ * `service_tier: "priority"`. A request asks for it only when the model
+ * binding's `fastModeAllowed` and the session `/fast` switch are both on.
+ */
+export const FAST_MODE_SERVICE_TIER = "priority";
+
+/** `subscription` spends ChatGPT plan limits; `usage` is billed per token. */
+export type FastModeBilling = "subscription" | "usage";
+
+/** Stored apiStyle and resolved wire spellings of the wires that carry the tier. */
+const FAST_MODE_BILLING: Record<string, FastModeBilling> = {
+  openai_codex_responses: "subscription",
+  "openai-codex-responses": "subscription",
+  responses: "usage",
+  "openai-responses": "usage",
+};
+
+/** How the tier is paid for on this apiStyle or wire, or undefined when it cannot carry it. */
+export function fastModeBillingOn(api: string | undefined | null): FastModeBilling | undefined {
+  return FAST_MODE_BILLING[(api ?? "").trim().toLowerCase()];
+}
+
+export function fastModeSupportedOn(api: string | undefined | null): boolean {
+  return fastModeBillingOn(api) !== undefined;
+}
diff --git a/packages/shared/src/index.ts b/packages/shared/src/index.ts
index 9ad014dc88..1e65c72bca 100644
--- a/packages/shared/src/index.ts
+++ b/packages/shared/src/index.ts
@@ -68,6 +68,7 @@ export * from "./session-presentation.js";
 export * from "./session-outcomes.js";
 export * from "./tray-sessions.js";
 export * from "./window-chrome.js";
+export * from "./fast-mode.js";
 export * from "./native-web-search.js";
 export * from "./native-web-search-transport.js";
 export * from "./header-value.js";
diff --git a/packages/shared/src/model-config-import.ts b/packages/shared/src/model-config-import.ts
index 02c30e4456..88fd99784f 100644
--- a/packages/shared/src/model-config-import.ts
+++ b/packages/shared/src/model-config-import.ts
@@ -839,6 +839,9 @@ function bindingFromGenericModel(
     ...(record?.nativeWebSearch === true || record?.native_web_search === true
       ? { nativeWebSearch: true }
       : {}),
+    ...(record?.fastModeAllowed === true || record?.fast_mode_allowed === true
+      ? { fastModeAllowed: true }
+      : {}),
   };
 }
 
diff --git a/packages/shared/src/provider-sync.ts b/packages/shared/src/provider-sync.ts
index 60a9fb1ead..467f1064f1 100644
--- a/packages/shared/src/provider-sync.ts
+++ b/packages/shared/src/provider-sync.ts
@@ -23,7 +23,7 @@ const API_STYLES = new Set(["auto", "chat_completions", "responses", "anthropic_
 const EXCLUDED_VENDORS = new Set(["claude-cli", "codex", "openai-codex", "github-copilot", "google-antigravity", "google-gemini-cli", "ollama", "lmstudio", "lm-studio", "local"]);
 const RESERVED_HEADERS = new Set(["authorization", "proxy-authorization", "x-api-key", "api-key", "x-goog-api-key", "cookie", "set-cookie", "host", "content-type", "content-length", "connection", "transfer-encoding", "te", "trailer", "upgrade", "keep-alive", "chatgpt-account-id", "x-opencode-session"]);
 const INPUT_FIELDS = ["name", "vendorKey", "type", "protocol", "baseUrl", "authKind", "models", "defaultModelId", "secretValue", "apiStyle", "headers", "supportsReasoning", "supportedThinkingLevels", "contextWindow", "maxOutputTokens", "temperature"];
-const MODEL_FIELDS = ["id", "alias", "contextWindow", "contextWindowSource", "maxTokens", "maxTokensSource", "thinkingLevels", "defaultThinkingLevel", "thinkingProtocol", "supportsImages", "supportsDocuments", "availableForSubagents", "nativeWebSearch"];
+const MODEL_FIELDS = ["id", "alias", "contextWindow", "contextWindowSource", "maxTokens", "maxTokensSource", "thinkingLevels", "defaultThinkingLevel", "thinkingProtocol", "supportsImages", "supportsDocuments", "availableForSubagents", "nativeWebSearch", "fastModeAllowed"];
 
 function invalid(): never {
   // Never include input values, unknown field names, parser diagnostics or causes.
@@ -85,7 +85,7 @@ function validateModel(value: unknown): ModelBinding {
   for (const key of ["supportsImages", "supportsDocuments"] as const) {
     if (value[key] !== undefined && value[key] !== null && typeof value[key] !== "boolean") invalid();
   }
-  for (const key of ["availableForSubagents", "nativeWebSearch"] as const) {
+  for (const key of ["availableForSubagents", "nativeWebSearch", "fastModeAllowed"] as const) {
     if (value[key] !== undefined && typeof value[key] !== "boolean") invalid();
   }
   return value as ModelBinding;
diff --git a/packages/shared/src/types/models.ts b/packages/shared/src/types/models.ts
index 47d436f6b6..856c94cc8d 100644
--- a/packages/shared/src/types/models.ts
+++ b/packages/shared/src/types/models.ts
@@ -254,6 +254,12 @@ export type ModelBinding = {
    * knowledge of the endpoint is the only source.
    */
   nativeWebSearch?: boolean;
+  /**
+   * Opt-in that lets the session `/fast` switch ask OpenAI for the priority
+   * service tier on this model (ADR 0332). Absent/false keeps the default
+   * tier; only Responses wires carry it (`fastModeSupportedOn`).
+   */
+  fastModeAllowed?: boolean;
 };
 
 export const MODEL_MODALITIES = ["text", "image", "audio", "video", "pdf"] as const;
diff --git a/packages/shared/src/types/sessions.ts b/packages/shared/src/types/sessions.ts
index a767955973..69ff8a78da 100644
--- a/packages/shared/src/types/sessions.ts
+++ b/packages/shared/src/types/sessions.ts
@@ -57,6 +57,12 @@ export type SessionSummary = {
   thinkingLevel: SessionThinkingLevel;
   /** Per-session permission mode; `inherit` follows the global default (D115). */
   permissionMode: PermissionMode;
+  /**
+   * Session `/fast` switch (ADR 0332). Requests use the priority tier only
+   * when the selected model also allows it. Older hosts and remote sessions
+   * omit it, meaning off.
+   */
+  fastMode?: boolean;
   /** Effective capability for this session's exact provider/model pair. */
   supportsReasoning?: boolean;
   /** Effective image-input capability for this session's exact model. */

From b4f958c4b74502be5f0ede923e7915b0c6fa1158 Mon Sep 17 00:00:00 2001
From: zszz3 <91608029+zszz3@users.noreply.github.com>
Date: Sun, 18 Oct 2026 12:31:07 +0800
Subject: [PATCH 2/3] refactor(agent-runtime): request Fast mode through
 pi-ai's serviceTier

pi-ai's Responses and Codex adapters already accept `serviceTier`, send
it as `service_tier`, and price the tier the response reports. Only
`streamSimple` dropped the option while building its base options, which
is why the first version wrote the field through a payload hook.

Forward `serviceTier` in the existing pi-ai patch and pass the native
option on main turns instead. The request bytes are unchanged, and the
Codex adapter can now price a priority request even when the backend
echoes the default tier.
---
 docs/adr/0332-openai-fast-mode.md             | 15 +++++----
 .../06-pi-runtime-dependency-boundary.md      |  3 +-
 .../06-pi-runtime-dependency-boundary.md      |  3 +-
 packages/agent-runtime/src/fast-mode.test.ts  | 23 ++++---------
 packages/agent-runtime/src/fast-mode.ts       | 29 ++++------------
 packages/agent-runtime/src/runtime.ts         |  1 +
 patches/@earendil-works__pi-ai@1.1.0.patch    | 33 ++++++++++++++-----
 pnpm-lock.yaml                                | 14 ++++----
 8 files changed, 58 insertions(+), 63 deletions(-)

diff --git a/docs/adr/0332-openai-fast-mode.md b/docs/adr/0332-openai-fast-mode.md
index 1c1a47cec2..074215f209 100644
--- a/docs/adr/0332-openai-fast-mode.md
+++ b/docs/adr/0332-openai-fast-mode.md
@@ -14,8 +14,9 @@ OpenAI serves a faster priority tier that Codex CLI and the OpenAI API call
 Fast mode. On the wire it is the Responses body field `service_tier:
 "priority"`; the Codex backend rejects `"fast"`. The tier costs more: a ChatGPT
 subscription spends its plan limits faster, and the OpenAI API bills a higher
-per-token price. PI-Desktop never asked for a tier, and pi-ai's `streamSimple`
-does not forward one.
+per-token price. PI-Desktop never asked for a tier. pi-ai's Responses and
+Codex adapters already accept `serviceTier` and price the tier the response
+reports, but `streamSimple` dropped the option while building its base options.
 
 The cost is a decision about one conversation, while whether an endpoint
 honors the field is something only the user knows for a relay. Issue #1657 asks
@@ -42,10 +43,12 @@ in model settings.
    the opt-in.
 4. **The launcher resolves; the runtime only stamps the wire.** Session launch
    sends `fastMode: true` to the sidecar only when the session switch is on and
-   the user's stored binding allows it. The runtime chains a payload hook on
-   main turns that adds `service_tier: "priority"` when the resolved wire is
-   `openai-codex-responses` or `openai-responses`. Subagents, compaction, and
-   one-shot plugin requests build their own options and keep the default tier.
+   the user's stored binding allows it. The runtime passes pi-ai's native
+   `serviceTier: "priority"` on main turns whose resolved wire is
+   `openai-codex-responses` or `openai-responses`; the existing pi-ai patch
+   forwards `serviceTier` through `streamSimple`, so the adapters also apply
+   their own tier pricing. Subagents, compaction, and one-shot plugin requests
+   build their own options and keep the default tier.
 
 ## Alternatives
 
diff --git a/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md b/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
index f91008399c..42b3ea1b67 100644
--- a/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
+++ b/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
@@ -19,7 +19,8 @@ in `docs/project/pi-110-adoption.md`.
   prompt-template expansion, and session-context projection.
 - `@earendil-works/pi-ai` supplies provider transports, auth, message types,
   request retries, and token estimation. Hosted-search request and estimation
-  behavior remains covered by PI-Desktop's pi-ai patch and contract tests.
+  behavior, and forwarding `serviceTier` through `streamSimple` for Fast mode
+  (ADR 0331), remain covered by PI-Desktop's pi-ai patch and contract tests.
 - Rust `host-core` remains the sole owner of SQLite and authoritative durable
   session state. The Node sidecar asks it to persist checkpoints and tool
   outcomes through the existing RPC contract.
diff --git a/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md b/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
index 006c7af228..c59f1582af 100644
--- a/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
+++ b/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
@@ -12,7 +12,8 @@ JavaScript 包。上游包演进时，必须保持这一职责划分清晰。
 - `@earendil-works/pi-agent-core` 提供 Agent 循环，以及稳定的 agent、event、tool
   类型。Agent Runtime 自行维护桌面专用的上下文投影、压缩和 prompt 模板展开。
 - `@earendil-works/pi-ai` 提供提供商传输、认证、消息类型、请求重试和 token 估算。
-  Hosted search 请求与估算行为继续由 PI-Desktop 的 pi-ai 补丁和契约测试覆盖。
+  Hosted search 请求与估算行为，以及为 Fast 模式让 `streamSimple` 转发 `serviceTier`
+  （ADR 0331），继续由 PI-Desktop 的 pi-ai 补丁和契约测试覆盖。
 - Rust `host-core` 仍然是 SQLite 和权威持久会话状态的唯一所有者。Node sidecar
   通过现有 RPC 契约请求持久化检查点和工具结果。
 - `@earendil-works/pi-coding-agent` 是过渡期依赖。生产代码仅在原生 Pi 会话续接及
diff --git a/packages/agent-runtime/src/fast-mode.test.ts b/packages/agent-runtime/src/fast-mode.test.ts
index 0bf505f499..8a395d2128 100644
--- a/packages/agent-runtime/src/fast-mode.test.ts
+++ b/packages/agent-runtime/src/fast-mode.test.ts
@@ -34,12 +34,12 @@ function capture() {
 const context = { messages: [{ role: "user", content: "hi", timestamp: 1 }] } as never;
 
 describe("withFastModeOptions", () => {
-  it("sends the priority tier on the Responses wire when allowed and switched on", async () => {
+  it("sends the priority tier through streamSimple on the Responses wire", async () => {
     const { bodies, options } = capture();
     await responsesStreamSimple(
       responsesModel,
       context,
-      withFastModeOptions(options, true),
+      withFastModeOptions(options, "openai-responses", true),
     ).result();
     expect(bodies).toHaveLength(1);
     expect(bodies[0].service_tier).toBe("priority");
@@ -47,7 +47,7 @@ describe("withFastModeOptions", () => {
 
   it("keeps the default tier when Fast mode is off", async () => {
     const { bodies, options } = capture();
-    const hooked = withFastModeOptions(options, false);
+    const hooked = withFastModeOptions(options, "openai-responses", false);
     expect(hooked).toBe(options);
     await responsesStreamSimple(responsesModel, context, hooked).result();
     expect(bodies[0].service_tier).toBeUndefined();
@@ -56,20 +56,9 @@ describe("withFastModeOptions", () => {
   it("leaves a Chat Completions request untouched even when switched on", async () => {
     const { bodies, options } = capture();
     const model = { ...responsesModel, api: "openai-completions" } as Model<"openai-completions">;
-    await completionsStreamSimple(
-      model,
-      context,
-      withFastModeOptions(options, true),
-    ).result();
+    const hooked = withFastModeOptions(options, model.api, true);
+    expect(hooked).toBe(options);
+    await completionsStreamSimple(model, context, hooked).result();
     expect(bodies[0].service_tier).toBeUndefined();
   });
-
-  it("chains after an earlier payload hook and keeps its replacement", async () => {
-    const options: SimpleStreamOptions = {
-      onPayload: async (payload) => ({ ...(payload as object), prompt_cache_key: "kept" }),
-    };
-    const hooked = withFastModeOptions(options, true);
-    const payload = await hooked.onPayload?.({ model: "gpt-test" }, responsesModel);
-    expect(payload).toEqual({ model: "gpt-test", prompt_cache_key: "kept", service_tier: "priority" });
-  });
 });
diff --git a/packages/agent-runtime/src/fast-mode.ts b/packages/agent-runtime/src/fast-mode.ts
index 893bc58a16..4b45d5812e 100644
--- a/packages/agent-runtime/src/fast-mode.ts
+++ b/packages/agent-runtime/src/fast-mode.ts
@@ -1,34 +1,17 @@
 /**
- * OpenAI Fast mode on the request body (ADR 0332).
- *
- * pi-ai's `streamSimple` does not forward a service tier, so the priority tier
- * is set through the payload hook, the same channel the session-affinity body
- * key uses. The launcher has already combined the session `/fast` switch with
- * the model binding's opt-in; the hook still reads the wire of the model being
- * called, so a request that does not travel on a Responses wire keeps its bytes.
+ * OpenAI Fast mode (ADR 0332). The launcher has already combined the session
+ * `/fast` switch with the model binding's opt-in; pi-ai's Responses adapters
+ * send `serviceTier` as `service_tier` and price the tier the response reports.
  */
 
 import type { SimpleStreamOptions } from "@earendil-works/pi-ai";
 import { FAST_MODE_SERVICE_TIER, fastModeSupportedOn } from "@pi-desktop/shared";
 
-/** Chain a payload hook that asks for the priority tier when Fast mode is on. */
+/** Ask for the priority tier on a wire that carries it; other wires keep their options. */
 export function withFastModeOptions(
   options: SimpleStreamOptions,
+  api: string,
   fastMode: boolean,
 ): SimpleStreamOptions {
-  if (fastMode !== true) return options;
-  const previous = options.onPayload;
-  return {
-    ...options,
-    onPayload: async (payload, model) => {
-      const replacement = await previous?.(payload, model);
-      const base = replacement === undefined ? payload : replacement;
-      if (!isRecord(base) || !fastModeSupportedOn(model.api)) return replacement;
-      return { ...base, service_tier: FAST_MODE_SERVICE_TIER };
-    },
-  };
-}
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return typeof value === "object" && value !== null && !Array.isArray(value);
+  return fastMode && fastModeSupportedOn(api) ? { ...options, serviceTier: FAST_MODE_SERVICE_TIER } : options;
 }
diff --git a/packages/agent-runtime/src/runtime.ts b/packages/agent-runtime/src/runtime.ts
index 5c576855a1..cc40c6c273 100644
--- a/packages/agent-runtime/src/runtime.ts
+++ b/packages/agent-runtime/src/runtime.ts
@@ -2203,6 +2203,7 @@ Do not invent objections or turn speculative risks into blockers. Stop when the
               sessionId: this.sessionId,
             },
           ),
+          m.api,
           this.fastMode,
         );
         // The watchdog must be able to *stop* what it abandons. It wraps the
diff --git a/patches/@earendil-works__pi-ai@1.1.0.patch b/patches/@earendil-works__pi-ai@1.1.0.patch
index 4ab80d45d0..caeea4f842 100644
--- a/patches/@earendil-works__pi-ai@1.1.0.patch
+++ b/patches/@earendil-works__pi-ai@1.1.0.patch
@@ -745,7 +745,7 @@ index 2a6b4e4..9e44070 100644
      const samplingParams = resolveSamplingParams(model, reasoningEffort ?? "off", options?.samplingParams);
      if (samplingParams) {
 diff --git a/dist/api/simple-options.js b/dist/api/simple-options.js
-index 4e36f35..c817c2e 100644
+index 4e36f35..fd9225e 100644
 --- a/dist/api/simple-options.js
 +++ b/dist/api/simple-options.js
 @@ -5,7 +5,7 @@ const MIN_MAX_TOKENS = 1;
@@ -757,6 +757,14 @@ index 4e36f35..c817c2e 100644
      return Math.min(maxTokens, Math.max(MIN_MAX_TOKENS, available));
  }
  export function resolveSamplingParams(model, thinkingLevel, requestParams) {
+@@ -30,6 +30,7 @@ export function buildBaseOptions(model, context, options, apiKey) {
+         sessionId: options?.sessionId,
+         headers: options?.headers,
+         onPayload: options?.onPayload,
++        serviceTier: options?.serviceTier,
+         onResponse: options?.onResponse,
+         onProviderStreamEvent: options?.onProviderStreamEvent,
+         timeoutMs: options?.timeoutMs,
 diff --git a/dist/auth/oauth/anthropic.js b/dist/auth/oauth/anthropic.js
 index 0823512..713d38a 100644
 --- a/dist/auth/oauth/anthropic.js
@@ -947,10 +955,19 @@ index 1443bb5..5fc0625 100644
 -{"openai-completions":{"chat:deepseek-flash":{"id":"deepseek-flash","name":"DeepSeek V4.1 Flash","api":"openai-completions","baseUrl":"https://api.deepseek.com","provider":"deepseek","reasoning":true,"thinkingLevelMap":{"minimal":null,"low":"low","medium":null,"high":"high","max":"max"},"input":["text","image"],"cost":{"input":0.3,"output":1.2,"cacheRead":0.006,"cacheWrite":0},"contextWindow":1000000,"maxTokens":384000,"compat":{"supportsStore":false,"supportsDeveloperRole":false,"maxTokensField":"max_tokens","requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","supportsStrictMode":true},"inputLimits":{"images":{"resize":{"maxWidth":2000,"maxHeight":2000,"maxBytes":4718592,"jpegQuality":80}}},"type":"chat"},"chat:deepseek-v4-pro":{"id":"deepseek-v4-pro","name":"DeepSeek V4 Pro","api":"openai-completions","baseUrl":"https://api.deepseek.com","provider":"deepseek","reasoning":true,"input":["text"],"cost":{"input":1.32,"output":3.96,"cacheRead":0.044,"cacheWrite":0},"contextWindow":1000000,"maxTokens":384000,"compat":{"supportsStore":false,"supportsDeveloperRole":false,"maxTokensField":"max_tokens","requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","supportsStrictMode":true,"supportsMidConvoSystemMessages":true},"thinkingLevelMap":{"minimal":null,"low":null,"medium":null,"high":"high","max":"max"},"type":"chat"}}}
 +{"openai-completions":{"chat:deepseek-flash":{"id":"deepseek-flash","name":"DeepSeek V4.1 Flash","api":"openai-completions","baseUrl":"https://api.deepseek.com","provider":"deepseek","reasoning":true,"thinkingLevelMap":{"minimal":null,"low":"low","medium":null,"high":"high","max":"max"},"input":["text","image"],"cost":{"input":0.3,"output":1.2,"cacheRead":0.006,"cacheWrite":0},"contextWindow":1000000,"maxTokens":384000,"compat":{"supportsStore":false,"supportsDeveloperRole":false,"maxTokensField":"max_tokens","requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","supportsStrictMode":true,"supportsMidConvoSystemMessages":true},"inputLimits":{"images":{"resize":{"maxWidth":2000,"maxHeight":2000,"maxBytes":4718592,"jpegQuality":80}}},"type":"chat"},"chat:deepseek-v4-pro":{"id":"deepseek-v4-pro","name":"DeepSeek V4 Pro","api":"openai-completions","baseUrl":"https://api.deepseek.com","provider":"deepseek","reasoning":true,"input":["text"],"cost":{"input":1.32,"output":3.96,"cacheRead":0.044,"cacheWrite":0},"contextWindow":1000000,"maxTokens":384000,"compat":{"supportsStore":false,"supportsDeveloperRole":false,"maxTokensField":"max_tokens","requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","supportsStrictMode":true,"supportsMidConvoSystemMessages":true},"thinkingLevelMap":{"minimal":null,"low":null,"medium":null,"high":"high","max":"max"},"type":"chat"}}}
 diff --git a/dist/types.d.ts b/dist/types.d.ts
-index b01603b..39a6d0a 100644
+index b01603b..432fbf7 100644
 --- a/dist/types.d.ts
 +++ b/dist/types.d.ts
-@@ -288,6 +288,30 @@ export interface ToolCall {
+@@ -251,6 +251,8 @@ export interface SimpleStreamOptions extends StreamOptions {
+     };
+     /** Custom token budgets for thinking levels (token-based providers only) */
+     thinkingBudgets?: ThinkingBudgets;
++    /** OpenAI Responses service tier; adapters without the field ignore it. */
++    serviceTier?: "auto" | "default" | "flex" | "scale" | "priority";
+ }
+ export type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (model: Model<TApi>, context: TranscriptContext, options?: TOptions) => AssistantMessageEventStream;
+ export type ImagesFunction<TOptions extends ImagesOptions = ImagesOptions> = (model: ImageModel<ImageApi>, context: ImagesContext, options?: TOptions) => Promise<AssistantImages>;
+@@ -288,6 +290,30 @@ export interface ToolCall {
      /** OpenAI Responses namespace for calls to dynamically loaded or namespaced tools. */
      namespace?: string;
  }
@@ -981,7 +998,7 @@ index b01603b..39a6d0a 100644
  export interface Usage {
      input: number;
      output: number;
-@@ -373,7 +397,7 @@ export interface UserMessage {
+@@ -373,7 +399,7 @@ export interface UserMessage {
  }
  export interface AssistantMessage {
      role: "assistant";
@@ -990,7 +1007,7 @@ index b01603b..39a6d0a 100644
      api: Api;
      provider: ProviderId;
      model: string;
-@@ -388,6 +412,8 @@ export interface AssistantMessage {
+@@ -388,6 +414,8 @@ export interface AssistantMessage {
      stopReason: StopReason;
      deferred?: DeferredHandle;
      errorMessage?: string;
@@ -999,7 +1016,7 @@ index b01603b..39a6d0a 100644
      rawStopReason?: string;
      /**
       * Provider indication of whether the model explicitly ended its turn.
-@@ -623,6 +649,10 @@ export type AssistantMessageEvent = {
+@@ -623,6 +651,10 @@ export type AssistantMessageEvent = {
      contentIndex: number;
      toolCall: ToolCall;
      partial: AssistantMessage;
@@ -1010,7 +1027,7 @@ index b01603b..39a6d0a 100644
  } | {
      type: "done";
      reason: Extract<StopReason, "stop" | "length" | "toolUse" | "deferred">;
-@@ -657,6 +687,12 @@ export interface OpenAICompletionsCompat {
+@@ -657,6 +689,12 @@ export interface OpenAICompletionsCompat {
      requiresThinkingAsText?: boolean;
      /** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */
      requiresReasoningContentOnAssistantMessages?: boolean;
@@ -1023,7 +1040,7 @@ index b01603b..39a6d0a 100644
      /** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "deepseek" uses thinking: { type } plus reasoning_effort when supported, "together" uses reasoning: { enabled } plus reasoning_effort when supported, "baseten" uses configurable chat_template_args plus reasoning_effort when supported, "zai" uses thinking: { type }, "qwen" uses top-level enable_thinking: boolean, "qwen-chat-template" uses chat_template_kwargs.enable_thinking and preserve_thinking, "chat-template" uses configurable chat_template_kwargs, "string-thinking" uses top-level thinking: string, and "ant-ling" uses reasoning: { effort } only when the mapped effort is non-null. Default: "openai". */
      thinkingFormat?: "openai" | "openrouter" | "deepseek" | "together" | "baseten" | "zai" | "qwen" | "chat-template" | "qwen-chat-template" | "string-thinking" | "ant-ling";
      /** Kwargs to send as `chat_template_kwargs` when `thinkingFormat` is `chat-template`. Use `{ "$var": "thinking.enabled" }`, `{ "$var": "thinking.effort" }`, or `{ "$var": "thinking.budget" }` for pi-controlled thinking values. */
-@@ -935,6 +971,8 @@ export interface Model<TApi extends Api> extends BaseModel<TApi> {
+@@ -935,6 +973,8 @@ export interface Model<TApi extends Api> extends BaseModel<TApi> {
       */
      type?: "chat";
      reasoning: boolean;
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 1ac3e20d51..43bcc8d0ed 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -206,7 +206,7 @@ overrides:
 
 patchedDependencies:
   '@earendil-works/pi-agent-core@1.1.0': a2c96619d0f9a1d643a6c6a555ba6a7fb533147336c181f7d3b7f847be33a6eb
-  '@earendil-works/pi-ai@1.1.0': 2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705
+  '@earendil-works/pi-ai@1.1.0': f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b
   '@earendil-works/pi-coding-agent@1.1.0': 73b8776451352deeaa307306b1aaa3a22e2c2cfd6e06c802fcd390409912f60a
 
 importers:
@@ -234,7 +234,7 @@ importers:
     devDependencies:
       '@earendil-works/pi-ai':
         specifier: 1.1.0
-        version: 1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@6.29.0)(ws@8.21.1)(zod@4.4.3)
+        version: 1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@6.29.0)(ws@8.21.1)(zod@4.4.3)
       '@earendil-works/pi-mcp':
         specifier: 1.1.0
         version: 1.1.0
@@ -429,7 +429,7 @@ importers:
         version: 1.1.0(patch_hash=a2c96619d0f9a1d643a6c6a555ba6a7fb533147336c181f7d3b7f847be33a6eb)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
       '@earendil-works/pi-ai':
         specifier: 1.1.0
-        version: 1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
+        version: 1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
       '@earendil-works/pi-coding-agent':
         specifier: 1.1.0
         version: 1.1.0(patch_hash=73b8776451352deeaa307306b1aaa3a22e2c2cfd6e06c802fcd390409912f60a)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(ws@8.21.1)(zod@4.4.3)
@@ -5235,7 +5235,7 @@ snapshots:
 
   '@earendil-works/pi-agent-core@1.1.0(patch_hash=a2c96619d0f9a1d643a6c6a555ba6a7fb533147336c181f7d3b7f847be33a6eb)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)':
     dependencies:
-      '@earendil-works/pi-ai': 1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
+      '@earendil-works/pi-ai': 1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
       typebox: 1.3.27
     transitivePeerDependencies:
       - '@aws-sdk/credential-provider-node'
@@ -5250,7 +5250,7 @@ snapshots:
       - ws
       - zod
 
-  '@earendil-works/pi-ai@1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@6.29.0)(ws@8.21.1)(zod@4.4.3)':
+  '@earendil-works/pi-ai@1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@6.29.0)(ws@8.21.1)(zod@4.4.3)':
     dependencies:
       '@anthropic-ai/sdk': 0.129.0(zod@4.4.3)
       '@aws-sdk/client-bedrock-runtime': 3.1127.0
@@ -5275,7 +5275,7 @@ snapshots:
       - ws
       - zod
 
-  '@earendil-works/pi-ai@1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)':
+  '@earendil-works/pi-ai@1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)':
     dependencies:
       '@anthropic-ai/sdk': 0.129.0(zod@4.4.3)
       '@aws-sdk/client-bedrock-runtime': 3.1127.0
@@ -5308,7 +5308,7 @@ snapshots:
     dependencies:
       '@earendil-works/chord': 1.1.0
       '@earendil-works/pi-agent-core': 1.1.0(patch_hash=a2c96619d0f9a1d643a6c6a555ba6a7fb533147336c181f7d3b7f847be33a6eb)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
-      '@earendil-works/pi-ai': 1.1.0(patch_hash=2aa7a30fca930a08ebdff65f82d2fc47aeb1cd941d34aae2e50e13b309ef2705)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
+      '@earendil-works/pi-ai': 1.1.0(patch_hash=f4ef44e75753e4d966c253f28d09a55924c26bb8a79bde1628ee66b2d8039b0b)(@aws-sdk/credential-provider-node@3.972.83)(@smithy/signature-v4@5.7.3)(supports-color@7.2.0)(undici@8.10.2)(ws@8.21.1)(zod@4.4.3)
       '@earendil-works/pi-codemode': 1.1.0
       '@earendil-works/pi-mcp': 1.1.0
       '@earendil-works/pi-tui': 1.1.0

From 2289727170815541d7fcddbc5f38ff1e92c33f03 Mon Sep 17 00:00:00 2001
From: zszz3 <91608029+zszz3@users.noreply.github.com>
Date: Mon, 19 Oct 2026 11:26:24 +0800
Subject: [PATCH 3/3] docs(adr): renumber Fast mode to ADR 0332 after rebase

Upstream main took ADR 0331 for plugin session follow and work-panel
reveal while this branch was in review. Renumber the Fast mode record
and its references so the two decisions keep distinct numbers.
---
 docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md  | 2 +-
 .../spec/02-architecture/06-pi-runtime-dependency-boundary.md   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md b/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
index 42b3ea1b67..c53fb0243f 100644
--- a/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
+++ b/docs/spec/02-architecture/06-pi-runtime-dependency-boundary.md
@@ -20,7 +20,7 @@ in `docs/project/pi-110-adoption.md`.
 - `@earendil-works/pi-ai` supplies provider transports, auth, message types,
   request retries, and token estimation. Hosted-search request and estimation
   behavior, and forwarding `serviceTier` through `streamSimple` for Fast mode
-  (ADR 0331), remain covered by PI-Desktop's pi-ai patch and contract tests.
+  (ADR 0332), remain covered by PI-Desktop's pi-ai patch and contract tests.
 - Rust `host-core` remains the sole owner of SQLite and authoritative durable
   session state. The Node sidecar asks it to persist checkpoints and tool
   outcomes through the existing RPC contract.
diff --git a/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md b/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
index c59f1582af..f79163d034 100644
--- a/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
+++ b/docs/zh-CN/spec/02-architecture/06-pi-runtime-dependency-boundary.md
@@ -13,7 +13,7 @@ JavaScript 包。上游包演进时，必须保持这一职责划分清晰。
   类型。Agent Runtime 自行维护桌面专用的上下文投影、压缩和 prompt 模板展开。
 - `@earendil-works/pi-ai` 提供提供商传输、认证、消息类型、请求重试和 token 估算。
   Hosted search 请求与估算行为，以及为 Fast 模式让 `streamSimple` 转发 `serviceTier`
-  （ADR 0331），继续由 PI-Desktop 的 pi-ai 补丁和契约测试覆盖。
+  （ADR 0332），继续由 PI-Desktop 的 pi-ai 补丁和契约测试覆盖。
 - Rust `host-core` 仍然是 SQLite 和权威持久会话状态的唯一所有者。Node sidecar
   通过现有 RPC 契约请求持久化检查点和工具结果。
 - `@earendil-works/pi-coding-agent` 是过渡期依赖。生产代码仅在原生 Pi 会话续接及