FEATURE: Fast and Ultrafast service tiers for Pi models

apps/mobile/src/components/ComposerModelControl.tsx

@@ -28,7 +28,6 @@ export function ComposerModelControl(props: {
   const promptTraits = getPromptControlledTraits(descriptors, props.prompt);
   const summary = props.provider
     ? buildTraitsTriggerDisplay({
-        provider: props.provider,
         descriptors,
         primarySelectDescriptorId: promptTraits.primarySelectDescriptor?.id ?? null,
         ultrathinkPromptControlled: promptTraits.ultrathinkPromptControlled,

apps/web/src/components/chat/ComposerModelEffort.tsx

@@ -36,16 +36,12 @@ import {
 type SelectDescriptor = Extract<ProviderOptionDescriptor, { type: "select" }>;
 type BooleanDescriptor = Extract<ProviderOptionDescriptor, { type: "boolean" }>;
 
-/** Codex service tiers, Claude's Fast mode, and Ultracode sit in the footer, not the column. */
+/** Service tiers, Claude's Fast mode, and Ultracode sit in the footer, not the column. */
 function splitSpeed(
-  provider: TraitsSelectionInput["provider"],
   selects: ReadonlyArray<SelectDescriptor>,
   toggles: ReadonlyArray<BooleanDescriptor>,
 ) {
-  const speedSelect =
-    provider === "codex"
-      ? (selects.find((descriptor) => descriptor.id === "serviceTier") ?? null)
-      : null;
+  const speedSelect = selects.find((descriptor) => descriptor.id === "serviceTier") ?? null;
   const fastToggle = toggles.find((descriptor) => descriptor.id === "fastMode") ?? null;
   const ultracodeToggle = toggles.find((descriptor) => descriptor.id === "ultracode") ?? null;
   return {
@@ -68,11 +64,10 @@ type ColumnControl =
  * such as an agent's session mode, sit below as one row each.
  */
 function columnControls(
-  provider: TraitsSelectionInput["provider"],
   selects: ReadonlyArray<SelectDescriptor>,
   toggles: ReadonlyArray<BooleanDescriptor>,
 ) {
-  const { columnSelects, columnToggles } = splitSpeed(provider, selects, toggles);
+  const { columnSelects, columnToggles } = splitSpeed(selects, toggles);
   const controls: ReadonlyArray<ColumnControl> = [
     ...columnSelects.map((descriptor) => ({ kind: "select" as const, descriptor })),
     ...columnToggles.map((descriptor) => ({ kind: "boolean" as const, descriptor })),
@@ -100,7 +95,6 @@ export function describeComposerEffort(
       : null;
   const primary = visibility.primarySelectDescriptor;
   const display = buildTraitsTriggerDisplay({
-    provider: traits.provider,
     descriptors: visibility.descriptors,
     primarySelectDescriptorId: primary?.id ?? null,
     ultrathinkPromptControlled: visibility.ultrathinkPromptControlled,
@@ -109,8 +103,8 @@ export function describeComposerEffort(
   });
   // The same reasoning control the column lists, wherever the agent put it among its options.
   const reasoning =
-    columnControls(traits.provider, visibility.selectDescriptors, visibility.booleanDescriptors)
-      .effort?.descriptor ?? null;
+    columnControls(visibility.selectDescriptors, visibility.booleanDescriptors).effort
+      ?.descriptor ?? null;
   const ultrathink = visibility.ultrathinkPromptControlled && reasoning === primary;
   const glyph =
     reasoning?.type === "select"
@@ -253,11 +247,7 @@ function moveBetweenRows(event: KeyboardEvent<HTMLElement>) {
 
 function EffortLevels(props: { traits: TraitsSelectionInput }) {
   const traits = useTraitsSelection(props.traits);
-  const { effort, others } = columnControls(
-    props.traits.provider,
-    traits.selectDescriptors,
-    traits.booleanDescriptors,
-  );
+  const { effort, others } = columnControls(traits.selectDescriptors, traits.booleanDescriptors);
   if (traits.modelIsUnavailable) {
     return (
       <>
@@ -505,11 +495,7 @@ export function ComposerEffortColumn(props: { traits: TraitsSelectionInput | nul
   const visibility = props.traits ? getTraitsSectionVisibility(props.traits) : null;
   const controls =
     props.traits && visibility
-      ? columnControls(
-          props.traits.provider,
-          visibility.selectDescriptors,
-          visibility.booleanDescriptors,
-        )
+      ? columnControls(visibility.selectDescriptors, visibility.booleanDescriptors)
       : null;
   return (
     <>
@@ -573,7 +559,6 @@ function SpeedSegment(props: {
 function SpeedControl(props: { traits: TraitsSelectionInput }) {
   const traits = useTraitsSelection(props.traits);
   const { speedSelect, fastToggle } = splitSpeed(
-    props.traits.provider,
     traits.selectDescriptors,
     traits.booleanDescriptors,
   );
@@ -610,11 +595,7 @@ function SpeedControl(props: { traits: TraitsSelectionInput }) {
 
 function UltracodeControl(props: { traits: TraitsSelectionInput }) {
   const traits = useTraitsSelection(props.traits);
-  const { ultracodeToggle } = splitSpeed(
-    props.traits.provider,
-    traits.selectDescriptors,
-    traits.booleanDescriptors,
-  );
+  const { ultracodeToggle } = splitSpeed(traits.selectDescriptors, traits.booleanDescriptors);
   if (traits.modelIsUnavailable || !ultracodeToggle) return null;
   const on = ultracodeToggle.currentValue === true;
   return (

apps/web/src/components/chat/SubagentTooltipContent.tsx

@@ -84,19 +84,18 @@ export function SubagentTooltipContent(props: {
             descriptor.type === "boolean" &&
             typeof saved?.value === "boolean"
           ) {
-            return getTraitsSpeedDisplay(provider.driver, {
+            return getTraitsSpeedDisplay({
               ...descriptor,
               currentValue: saved.value,
             });
           }
           if (
-            provider.driver === "codex" &&
             descriptor.id === "serviceTier" &&
             descriptor.type === "select" &&
             typeof saved?.value === "string" &&
             descriptor.options.some((option) => option.id === saved.value)
           ) {
-            return getTraitsSpeedDisplay(provider.driver, {
+            return getTraitsSpeedDisplay({
               ...descriptor,
               currentValue: saved.value,
             });

apps/web/src/components/chat/TraitsPicker.tsx

@@ -495,7 +495,6 @@ export const TraitsPicker = memo(function TraitsPicker({
     return null;
   }
   const { label: triggerLabel } = buildTraitsTriggerDisplay({
-    provider: props.provider,
     descriptors: traits.descriptors,
     primarySelectDescriptorId: traits.primarySelectDescriptor?.id ?? null,
     ultrathinkPromptControlled: traits.ultrathinkPromptControlled,

apps/web/src/components/chat/TraitsSpeed.tsx

@@ -6,7 +6,6 @@ import { cn } from "~/lib/utils";
 import { ComposerControlIcon, type ComposerControlSize } from "./ComposerControl";
 
 export function getTraitsSpeedDisplay(
-  provider: ProviderDriverKind,
   descriptor: ProviderOptionDescriptor,
 ): { label: string; speedIcon: "fast" | "ultrafast" | null } | null {
   if (descriptor.id === "fastMode" && descriptor.type === "boolean") {
@@ -15,7 +14,7 @@ export function getTraitsSpeedDisplay(
       speedIcon: descriptor.currentValue === true ? "fast" : null,
     };
   }
-  if (provider !== "codex" || descriptor.id !== "serviceTier" || descriptor.type !== "select") {
+  if (descriptor.id !== "serviceTier" || descriptor.type !== "select") {
     return null;
   }
   const currentValue = getProviderOptionCurrentValue(descriptor);

docs/user/providers-pi.md

@@ -37,7 +37,10 @@ update it through that manager instead. For example, a Nix profile installation
 
 T3 Code discovers the models reported by Pi and exposes their supported thinking levels. The
 thinking picker marks the discovered active model's configured level as its default without
-overriding it. Other models keep their Pi defaults when no level is selected. Threads
+overriding it. Other models keep their Pi defaults when no level is selected. A model that lists
+`serviceTiers` (for example `[{ "id": "priority", "name": "Fast" }]`) in its Pi definition gets
+a speed picker with Standard plus those tiers. A chosen tier is sent as the request's
+`service_tier`; Standard sends none and leaves the endpoint's default. Threads
 use Pi's native session files for resume, rollback, and forks within the same Pi instance. Forks
 preserve the native conversation through the selected turn in the destination workspace.
 Switching providers uses portable conversation context. Extension

packages/client-runtime/src/modelTraits.test.ts

@@ -1,4 +1,4 @@
-import { ProviderDriverKind, type ProviderOptionDescriptor } from "@t3tools/contracts";
+import type { ProviderOptionDescriptor } from "@t3tools/contracts";
 import { describe, expect, it } from "vite-plus/test";
 
 import { buildTraitsTriggerDisplay, getDescriptorStringValue } from "./modelTraits.ts";
@@ -25,7 +25,6 @@ describe("buildTraitsTriggerDisplay", () => {
 
     expect(
       buildTraitsTriggerDisplay({
-        provider: ProviderDriverKind.make("pi"),
         descriptors: [descriptor],
         primarySelectDescriptorId: descriptor.id,
         ultrathinkPromptControlled: false,

packages/client-runtime/src/modelTraits.ts

@@ -1,8 +1,4 @@
-import type {
-  ModelSelection,
-  ProviderDriverKind,
-  ProviderOptionDescriptor,
-} from "@t3tools/contracts";
+import type { ModelSelection, ProviderOptionDescriptor } from "@t3tools/contracts";
 import {
   buildProviderOptionSelectionsFromDescriptors,
   getProviderOptionCurrentLabel,
@@ -110,7 +106,6 @@ export function getTraitSelectChange(
 }
 
 export function buildTraitsTriggerDisplay(input: {
-  provider: ProviderDriverKind;
   descriptors: ReadonlyArray<ProviderOptionDescriptor>;
   primarySelectDescriptorId: string | null;
   ultrathinkPromptControlled: boolean;
@@ -140,11 +135,7 @@ export function buildTraitsTriggerDisplay(input: {
       }
       continue;
     }
-    if (
-      input.provider === "codex" &&
-      descriptor.id === "serviceTier" &&
-      descriptor.type === "select"
-    ) {
+    if (descriptor.id === "serviceTier" && descriptor.type === "select") {
       const currentValue = getProviderOptionCurrentValue(descriptor);
       const fastTier = descriptor.options.find(({ label }) => label === "Fast");
       const ultrafastTier = descriptor.options.find(({ label }) => label === "Ultrafast");

packages/provider-pi/src/server/adapter.test.ts

@@ -53,6 +53,7 @@ import {
 } from "./adapter.ts";
 import { makePiRpcConnection, type PiRpcRecord } from "./rpc.ts";
 import { loadMcpBridge } from "./mcpBridge.testkit.ts";
+import { PI_T3_SERVICE_TIER_COMMAND } from "./mcpExtensionSource.ts";
 import { turnItemOutputText } from "../../../client-runtime/src/work-log/itemDetail.ts";
 
 const layerTest = Layer.mergeAll(
@@ -173,6 +174,12 @@ interface FakePi {
   readonly allRequests: () => ReadonlyArray<PiRpcRecord>;
   readonly promptThinkingLevels: ReadonlyArray<string>;
   readonly promptModels: ReadonlyArray<string>;
+  /** T3 extension's service tier as each prompt was sent; Pi resets it with each session. */
+  readonly promptServiceTiers: ReadonlyArray<string | null>;
+  /** Pi reloads extensions without the adapter seeing it, e.g. a user extension's ctx.newSession(). */
+  readonly reloadExtensions: () => void;
+  /** Pi before 0.99 acknowledges prompts without a disposition. */
+  readonly omitPromptDisposition: () => void;
   readonly queueStats: (data: unknown) => void;
   /** Pi before 0.84.4 answers clear_queue as an unknown command. */
   readonly setClearQueue: (
@@ -230,7 +237,11 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
   let thinkingLevel = "high";
   const promptThinkingLevels: Array<string> = [];
   const promptModels: Array<string> = [];
+  const promptServiceTiers: Array<string | null> = [];
+  let serviceTier: string | null = null;
+  let promptDisposition = true;
   let currentModel = "default";
+  const serviceTierCommand = `/${PI_T3_SERVICE_TIER_COMMAND}`;
 
   const emit = (record: PiRpcRecord, chunkBytes?: number) =>
     Effect.gen(function* () {
@@ -280,6 +291,10 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
         thinkingLevel = String(record["level"]);
         return base;
       case "prompt":
+        if (String(record["message"]).startsWith(serviceTierCommand)) {
+          serviceTier = String(record["message"]).slice(serviceTierCommand.length).trim() || null;
+          return promptDisposition ? { ...base, data: { disposition: "handled" } } : base;
+        }
         if (isCompacting) {
           return { ...base, success: false, error: PI_COMPACTING_PROMPT_ERROR };
         }
@@ -297,12 +312,16 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
       case "new_session": {
         const cancelled = vetoNewSession;
         vetoNewSession = false;
-        if (!cancelled) sessionFile = `/fake/new-${++sessionGeneration}.jsonl`;
+        if (!cancelled) {
+          sessionFile = `/fake/new-${++sessionGeneration}.jsonl`;
+          serviceTier = null;
+        }
         return { ...base, data: { cancelled } };
       }
       case "switch_session": {
         const cancelled = vetoSwitch;
         vetoSwitch = false;
+        if (!cancelled) serviceTier = null;
         return { ...base, data: { cancelled } };
       }
       case "get_entries":
@@ -312,6 +331,7 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
       case "get_session_stats":
         return { ...base, data: statsQueue.shift() ?? {} };
       case "fork":
+        serviceTier = null;
         return { ...base, data: { text: "Hello pi", cancelled: false } };
       case "clear_queue":
         return clearQueue === "unknown_command"
@@ -332,9 +352,13 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
         stdinBuffer = stdinBuffer.slice(newline + 1);
         if (line.length === 0) continue;
         const record = decodeJsonLine(line) as PiRpcRecord;
-        if (record["type"] === "prompt") {
+        if (
+          record["type"] === "prompt" &&
+          !String(record["message"]).startsWith(serviceTierCommand)
+        ) {
           promptThinkingLevels.push(thinkingLevel);
           promptModels.push(currentModel);
+          promptServiceTiers.push(serviceTier);
         }
         allRequests.push(record);
         yield* Queue.offer(requests, record);
@@ -395,6 +419,13 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
     },
     promptThinkingLevels,
     promptModels,
+    promptServiceTiers,
+    reloadExtensions: () => {
+      serviceTier = null;
+    },
+    omitPromptDisposition: () => {
+      promptDisposition = false;
+    },
     emit,
     takeRequest,
     queueEntries: (data) => entriesQueue.push(data),
@@ -2293,6 +2324,68 @@ describe("PiAdapterV2", () => {
       }).pipe(Effect.scoped, Effect.provide(layerTest)),
   );
 
+  it.effect("applies the selected service tier without changing thinking", () =>
+    Effect.gen(function* () {
+      const fake = yield* makeFakePi;
+      const { runtime, takeEvent } = yield* openRuntime(fake);
+      const providerThread = yield* runtime.ensureThread({
+        threadId: THREAD_ID,
+        modelSelection: modelSelection("default"),
+        runtimePolicy,
+      });
+      const turn = (text: string, serviceTier: string | undefined, ordinal: number) =>
+        Effect.gen(function* () {
+          yield* startTurn(
+            runtime,
+            providerThread,
+            "default",
+            [],
+            text,
+            {
+              instanceId: PI_INSTANCE_ID,
+              model: "columbia-codex/gpt-6.1-sol",
+              options: [
+                { id: "thinking", value: "high" },
+                ...(serviceTier === undefined ? [] : [{ id: "serviceTier", value: serviceTier }]),
+              ],
+            },
+            ordinal,
+          );
+          while (String((yield* fake.takeRequest("prompt"))["message"]).startsWith("/")) {}
+          yield* fake.emit({ type: "agent_start" });
+          yield* fake.emit({ type: "agent_end", messages: [], willRetry: false });
+          yield* fake.emit({ type: "agent_settled" });
+          yield* takeEvent((event) => event.type === "turn.terminal");
+        });
+
+      yield* turn("Fast", "priority", 1);
+      fake.reloadExtensions();
+      yield* turn("Still fast", "priority", 2);
+      yield* turn("Ultrafast", "ultrafast", 3);
+      fake.omitPromptDisposition();
+      yield* turn("Standard", "default", 4);
+      yield* turn("Unset", undefined, 5);
+
+      assert.deepEqual(fake.promptServiceTiers, ["priority", "priority", "ultrafast", null, null]);
+      assert.deepEqual(fake.promptThinkingLevels, ["high", "high", "high", "high", "high"]);
+      assert.deepEqual(
+        fake
+          .allRequests()
+          .flatMap((request) =>
+            String(request["message"]).startsWith(`/${PI_T3_SERVICE_TIER_COMMAND}`)
+              ? [request["message"]]
+              : [],
+          ),
+        [
+          `/${PI_T3_SERVICE_TIER_COMMAND} priority`,
+          `/${PI_T3_SERVICE_TIER_COMMAND} priority`,
+          `/${PI_T3_SERVICE_TIER_COMMAND} ultrafast`,
+          `/${PI_T3_SERVICE_TIER_COMMAND}`,
+        ],
+      );
+    }).pipe(Effect.scoped, Effect.provide(layerTest)),
+  );
+
   it.effect.each([false, true, "previous"] as const)(
     "resets applied thinking when returning to Pi default (failed first restore: %s)",
     (failFirstRestore) =>