@devries-cameron20 / t3code Mirror · read-only / commits / 225685a8d5
Fork 0

The best way to control Claude Code, Codex and more from any computer or phone

Mirrored from https://github.com/Camo67/t3code.git

Synced Nov 22, 02:00 UTC

feat(pi): Fast and Ultrafast service tiers for Pi models (#33177)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Ben Davis committed Nov 8, 2026 at 20:17 UTC 225685a8d5a620503d7713a8e48131f567df9f6b
16 files changed +319 -89
apps/mobile/src/components/ComposerModelControl.tsx
-1
@@ -28,7 +28,6 @@ export function ComposerModelControl(props: {
28 const promptTraits = getPromptControlledTraits(descriptors, props.prompt);
29 const summary = props.provider
30 ? buildTraitsTriggerDisplay({
31 - provider: props.provider,
31 descriptors,
32 primarySelectDescriptorId: promptTraits.primarySelectDescriptor?.id ?? null,
33 ultrathinkPromptControlled: promptTraits.ultrathinkPromptControlled,
apps/web/src/components/chat/ComposerModelEffort.tsx
+8 -27
@@ -36,16 +36,12 @@ import {
36 type SelectDescriptor = Extract<ProviderOptionDescriptor, { type: "select" }>;
37 type BooleanDescriptor = Extract<ProviderOptionDescriptor, { type: "boolean" }>;
38
39 -/** Codex service tiers, Claude's Fast mode, and Ultracode sit in the footer, not the column. */
39 +/** Service tiers, Claude's Fast mode, and Ultracode sit in the footer, not the column. */
40 function splitSpeed(
41 - provider: TraitsSelectionInput["provider"],
41 selects: ReadonlyArray<SelectDescriptor>,
42 toggles: ReadonlyArray<BooleanDescriptor>,
43 ) {
45 - const speedSelect =
46 - provider === "codex"
47 - ? (selects.find((descriptor) => descriptor.id === "serviceTier") ?? null)
48 - : null;
44 + const speedSelect = selects.find((descriptor) => descriptor.id === "serviceTier") ?? null;
45 const fastToggle = toggles.find((descriptor) => descriptor.id === "fastMode") ?? null;
46 const ultracodeToggle = toggles.find((descriptor) => descriptor.id === "ultracode") ?? null;
47 return {
@@ -68,11 +64,10 @@ type ColumnControl =
64 * such as an agent's session mode, sit below as one row each.
65 */
66 function columnControls(
71 - provider: TraitsSelectionInput["provider"],
67 selects: ReadonlyArray<SelectDescriptor>,
68 toggles: ReadonlyArray<BooleanDescriptor>,
69 ) {
75 - const { columnSelects, columnToggles } = splitSpeed(provider, selects, toggles);
70 + const { columnSelects, columnToggles } = splitSpeed(selects, toggles);
71 const controls: ReadonlyArray<ColumnControl> = [
72 ...columnSelects.map((descriptor) => ({ kind: "select" as const, descriptor })),
73 ...columnToggles.map((descriptor) => ({ kind: "boolean" as const, descriptor })),
@@ -100,7 +95,6 @@ export function describeComposerEffort(
95 : null;
96 const primary = visibility.primarySelectDescriptor;
97 const display = buildTraitsTriggerDisplay({
103 - provider: traits.provider,
98 descriptors: visibility.descriptors,
99 primarySelectDescriptorId: primary?.id ?? null,
100 ultrathinkPromptControlled: visibility.ultrathinkPromptControlled,
@@ -109,8 +103,8 @@ export function describeComposerEffort(
103 });
104 // The same reasoning control the column lists, wherever the agent put it among its options.
105 const reasoning =
112 - columnControls(traits.provider, visibility.selectDescriptors, visibility.booleanDescriptors)
113 - .effort?.descriptor ?? null;
106 + columnControls(visibility.selectDescriptors, visibility.booleanDescriptors).effort
107 + ?.descriptor ?? null;
108 const ultrathink = visibility.ultrathinkPromptControlled && reasoning === primary;
109 const glyph =
110 reasoning?.type === "select"
@@ -253,11 +247,7 @@ function moveBetweenRows(event: KeyboardEvent<HTMLElement>) {
247
248 function EffortLevels(props: { traits: TraitsSelectionInput }) {
249 const traits = useTraitsSelection(props.traits);
256 - const { effort, others } = columnControls(
257 - props.traits.provider,
258 - traits.selectDescriptors,
259 - traits.booleanDescriptors,
260 - );
250 + const { effort, others } = columnControls(traits.selectDescriptors, traits.booleanDescriptors);
251 if (traits.modelIsUnavailable) {
252 return (
253 <>
@@ -505,11 +495,7 @@ export function ComposerEffortColumn(props: { traits: TraitsSelectionInput | nul
495 const visibility = props.traits ? getTraitsSectionVisibility(props.traits) : null;
496 const controls =
497 props.traits && visibility
508 - ? columnControls(
509 - props.traits.provider,
510 - visibility.selectDescriptors,
511 - visibility.booleanDescriptors,
512 - )
498 + ? columnControls(visibility.selectDescriptors, visibility.booleanDescriptors)
499 : null;
500 return (
501 <>
@@ -573,7 +559,6 @@ function SpeedSegment(props: {
559 function SpeedControl(props: { traits: TraitsSelectionInput }) {
560 const traits = useTraitsSelection(props.traits);
561 const { speedSelect, fastToggle } = splitSpeed(
576 - props.traits.provider,
562 traits.selectDescriptors,
563 traits.booleanDescriptors,
564 );
@@ -610,11 +595,7 @@ function SpeedControl(props: { traits: TraitsSelectionInput }) {
595
596 function UltracodeControl(props: { traits: TraitsSelectionInput }) {
597 const traits = useTraitsSelection(props.traits);
613 - const { ultracodeToggle } = splitSpeed(
614 - props.traits.provider,
615 - traits.selectDescriptors,
616 - traits.booleanDescriptors,
617 - );
598 + const { ultracodeToggle } = splitSpeed(traits.selectDescriptors, traits.booleanDescriptors);
599 if (traits.modelIsUnavailable || !ultracodeToggle) return null;
600 const on = ultracodeToggle.currentValue === true;
601 return (
apps/web/src/components/chat/SubagentTooltipContent.tsx
+2 -3
@@ -84,19 +84,18 @@ export function SubagentTooltipContent(props: {
84 descriptor.type === "boolean" &&
85 typeof saved?.value === "boolean"
86 ) {
87 - return getTraitsSpeedDisplay(provider.driver, {
87 + return getTraitsSpeedDisplay({
88 ...descriptor,
89 currentValue: saved.value,
90 });
91 }
92 if (
93 - provider.driver === "codex" &&
93 descriptor.id === "serviceTier" &&
94 descriptor.type === "select" &&
95 typeof saved?.value === "string" &&
96 descriptor.options.some((option) => option.id === saved.value)
97 ) {
99 - return getTraitsSpeedDisplay(provider.driver, {
98 + return getTraitsSpeedDisplay({
99 ...descriptor,
100 currentValue: saved.value,
101 });
apps/web/src/components/chat/TraitsPicker.tsx
-1
@@ -495,7 +495,6 @@ export const TraitsPicker = memo(function TraitsPicker({
495 return null;
496 }
497 const { label: triggerLabel } = buildTraitsTriggerDisplay({
498 - provider: props.provider,
498 descriptors: traits.descriptors,
499 primarySelectDescriptorId: traits.primarySelectDescriptor?.id ?? null,
500 ultrathinkPromptControlled: traits.ultrathinkPromptControlled,
apps/web/src/components/chat/TraitsSpeed.tsx
+1 -2
@@ -6,7 +6,6 @@ import { cn } from "~/lib/utils";
6 import { ComposerControlIcon, type ComposerControlSize } from "./ComposerControl";
7
8 export function getTraitsSpeedDisplay(
9 - provider: ProviderDriverKind,
9 descriptor: ProviderOptionDescriptor,
10 ): { label: string; speedIcon: "fast" | "ultrafast" | null } | null {
11 if (descriptor.id === "fastMode" && descriptor.type === "boolean") {
@@ -15,7 +14,7 @@ export function getTraitsSpeedDisplay(
14 speedIcon: descriptor.currentValue === true ? "fast" : null,
15 };
16 }
18 - if (provider !== "codex" || descriptor.id !== "serviceTier" || descriptor.type !== "select") {
17 + if (descriptor.id !== "serviceTier" || descriptor.type !== "select") {
18 return null;
19 }
20 const currentValue = getProviderOptionCurrentValue(descriptor);
docs/user/providers-pi.md
+4 -1
@@ -37,7 +37,10 @@ update it through that manager instead. For example, a Nix profile installation
37
38 T3 Code discovers the models reported by Pi and exposes their supported thinking levels. The
39 thinking picker marks the discovered active model's configured level as its default without
40 -overriding it. Other models keep their Pi defaults when no level is selected. Threads
40 +overriding it. Other models keep their Pi defaults when no level is selected. A model that lists
41 +`serviceTiers` (for example `[{ "id": "priority", "name": "Fast" }]`) in its Pi definition gets
42 +a speed picker with Standard plus those tiers. A chosen tier is sent as the request's
43 +`service_tier`; Standard sends none and leaves the endpoint's default. Threads
44 use Pi's native session files for resume, rollback, and forks within the same Pi instance. Forks
45 preserve the native conversation through the selected turn in the destination workspace.
46 Switching providers uses portable conversation context. Extension
packages/client-runtime/src/modelTraits.test.ts
+1 -2
@@ -1,4 +1,4 @@
1 -import { ProviderDriverKind, type ProviderOptionDescriptor } from "@t3tools/contracts";
1 +import type { ProviderOptionDescriptor } from "@t3tools/contracts";
2 import { describe, expect, it } from "vite-plus/test";
3
4 import { buildTraitsTriggerDisplay, getDescriptorStringValue } from "./modelTraits.ts";
@@ -25,7 +25,6 @@ describe("buildTraitsTriggerDisplay", () => {
25
26 expect(
27 buildTraitsTriggerDisplay({
28 - provider: ProviderDriverKind.make("pi"),
28 descriptors: [descriptor],
29 primarySelectDescriptorId: descriptor.id,
30 ultrathinkPromptControlled: false,
packages/client-runtime/src/modelTraits.ts
+2 -11
@@ -1,8 +1,4 @@
1 -import type {
2 - ModelSelection,
3 - ProviderDriverKind,
4 - ProviderOptionDescriptor,
5 -} from "@t3tools/contracts";
1 +import type { ModelSelection, ProviderOptionDescriptor } from "@t3tools/contracts";
2 import {
3 buildProviderOptionSelectionsFromDescriptors,
4 getProviderOptionCurrentLabel,
@@ -110,7 +106,6 @@ export function getTraitSelectChange(
106 }
107
108 export function buildTraitsTriggerDisplay(input: {
113 - provider: ProviderDriverKind;
109 descriptors: ReadonlyArray<ProviderOptionDescriptor>;
110 primarySelectDescriptorId: string | null;
111 ultrathinkPromptControlled: boolean;
@@ -140,11 +135,7 @@ export function buildTraitsTriggerDisplay(input: {
135 }
136 continue;
137 }
143 - if (
144 - input.provider === "codex" &&
145 - descriptor.id === "serviceTier" &&
146 - descriptor.type === "select"
147 - ) {
138 + if (descriptor.id === "serviceTier" && descriptor.type === "select") {
139 const currentValue = getProviderOptionCurrentValue(descriptor);
140 const fastTier = descriptor.options.find(({ label }) => label === "Fast");
141 const ultrafastTier = descriptor.options.find(({ label }) => label === "Ultrafast");
packages/provider-pi/src/server/adapter.test.ts
+95 -2
@@ -53,6 +53,7 @@ import {
53 } from "./adapter.ts";
54 import { makePiRpcConnection, type PiRpcRecord } from "./rpc.ts";
55 import { loadMcpBridge } from "./mcpBridge.testkit.ts";
56 +import { PI_T3_SERVICE_TIER_COMMAND } from "./mcpExtensionSource.ts";
57 import { turnItemOutputText } from "../../../client-runtime/src/work-log/itemDetail.ts";
58
59 const layerTest = Layer.mergeAll(
@@ -173,6 +174,12 @@ interface FakePi {
174 readonly allRequests: () => ReadonlyArray<PiRpcRecord>;
175 readonly promptThinkingLevels: ReadonlyArray<string>;
176 readonly promptModels: ReadonlyArray<string>;
177 + /** T3 extension's service tier as each prompt was sent; Pi resets it with each session. */
178 + readonly promptServiceTiers: ReadonlyArray<string | null>;
179 + /** Pi reloads extensions without the adapter seeing it, e.g. a user extension's ctx.newSession(). */
180 + readonly reloadExtensions: () => void;
181 + /** Pi before 0.99 acknowledges prompts without a disposition. */
182 + readonly omitPromptDisposition: () => void;
183 readonly queueStats: (data: unknown) => void;
184 /** Pi before 0.84.4 answers clear_queue as an unknown command. */
185 readonly setClearQueue: (
@@ -230,7 +237,11 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
237 let thinkingLevel = "high";
238 const promptThinkingLevels: Array<string> = [];
239 const promptModels: Array<string> = [];
240 + const promptServiceTiers: Array<string | null> = [];
241 + let serviceTier: string | null = null;
242 + let promptDisposition = true;
243 let currentModel = "default";
244 + const serviceTierCommand = `/${PI_T3_SERVICE_TIER_COMMAND}`;
245
246 const emit = (record: PiRpcRecord, chunkBytes?: number) =>
247 Effect.gen(function* () {
@@ -280,6 +291,10 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
291 thinkingLevel = String(record["level"]);
292 return base;
293 case "prompt":
294 + if (String(record["message"]).startsWith(serviceTierCommand)) {
295 + serviceTier = String(record["message"]).slice(serviceTierCommand.length).trim() || null;
296 + return promptDisposition ? { ...base, data: { disposition: "handled" } } : base;
297 + }
298 if (isCompacting) {
299 return { ...base, success: false, error: PI_COMPACTING_PROMPT_ERROR };
300 }
@@ -297,12 +312,16 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
312 case "new_session": {
313 const cancelled = vetoNewSession;
314 vetoNewSession = false;
300 - if (!cancelled) sessionFile = `/fake/new-${++sessionGeneration}.jsonl`;
315 + if (!cancelled) {
316 + sessionFile = `/fake/new-${++sessionGeneration}.jsonl`;
317 + serviceTier = null;
318 + }
319 return { ...base, data: { cancelled } };
320 }
321 case "switch_session": {
322 const cancelled = vetoSwitch;
323 vetoSwitch = false;
324 + if (!cancelled) serviceTier = null;
325 return { ...base, data: { cancelled } };
326 }
327 case "get_entries":
@@ -312,6 +331,7 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
331 case "get_session_stats":
332 return { ...base, data: statsQueue.shift() ?? {} };
333 case "fork":
334 + serviceTier = null;
335 return { ...base, data: { text: "Hello pi", cancelled: false } };
336 case "clear_queue":
337 return clearQueue === "unknown_command"
@@ -332,9 +352,13 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
352 stdinBuffer = stdinBuffer.slice(newline + 1);
353 if (line.length === 0) continue;
354 const record = decodeJsonLine(line) as PiRpcRecord;
335 - if (record["type"] === "prompt") {
355 + if (
356 + record["type"] === "prompt" &&
357 + !String(record["message"]).startsWith(serviceTierCommand)
358 + ) {
359 promptThinkingLevels.push(thinkingLevel);
360 promptModels.push(currentModel);
361 + promptServiceTiers.push(serviceTier);
362 }
363 allRequests.push(record);
364 yield* Queue.offer(requests, record);
@@ -395,6 +419,13 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
419 },
420 promptThinkingLevels,
421 promptModels,
422 + promptServiceTiers,
423 + reloadExtensions: () => {
424 + serviceTier = null;
425 + },
426 + omitPromptDisposition: () => {
427 + promptDisposition = false;
428 + },
429 emit,
430 takeRequest,
431 queueEntries: (data) => entriesQueue.push(data),
@@ -2293,6 +2324,68 @@ describe("PiAdapterV2", () => {
2324 }).pipe(Effect.scoped, Effect.provide(layerTest)),
2325 );
2326
2327 + it.effect("applies the selected service tier without changing thinking", () =>
2328 + Effect.gen(function* () {
2329 + const fake = yield* makeFakePi;
2330 + const { runtime, takeEvent } = yield* openRuntime(fake);
2331 + const providerThread = yield* runtime.ensureThread({
2332 + threadId: THREAD_ID,
2333 + modelSelection: modelSelection("default"),
2334 + runtimePolicy,
2335 + });
2336 + const turn = (text: string, serviceTier: string | undefined, ordinal: number) =>
2337 + Effect.gen(function* () {
2338 + yield* startTurn(
2339 + runtime,
2340 + providerThread,
2341 + "default",
2342 + [],
2343 + text,
2344 + {
2345 + instanceId: PI_INSTANCE_ID,
2346 + model: "columbia-codex/gpt-6.1-sol",
2347 + options: [
2348 + { id: "thinking", value: "high" },
2349 + ...(serviceTier === undefined ? [] : [{ id: "serviceTier", value: serviceTier }]),
2350 + ],
2351 + },
2352 + ordinal,
2353 + );
2354 + while (String((yield* fake.takeRequest("prompt"))["message"]).startsWith("/")) {}
2355 + yield* fake.emit({ type: "agent_start" });
2356 + yield* fake.emit({ type: "agent_end", messages: [], willRetry: false });
2357 + yield* fake.emit({ type: "agent_settled" });
2358 + yield* takeEvent((event) => event.type === "turn.terminal");
2359 + });
2360 +
2361 + yield* turn("Fast", "priority", 1);
2362 + fake.reloadExtensions();
2363 + yield* turn("Still fast", "priority", 2);
2364 + yield* turn("Ultrafast", "ultrafast", 3);
2365 + fake.omitPromptDisposition();
2366 + yield* turn("Standard", "default", 4);
2367 + yield* turn("Unset", undefined, 5);
2368 +
2369 + assert.deepEqual(fake.promptServiceTiers, ["priority", "priority", "ultrafast", null, null]);
2370 + assert.deepEqual(fake.promptThinkingLevels, ["high", "high", "high", "high", "high"]);
2371 + assert.deepEqual(
2372 + fake
2373 + .allRequests()
2374 + .flatMap((request) =>
2375 + String(request["message"]).startsWith(`/${PI_T3_SERVICE_TIER_COMMAND}`)
2376 + ? [request["message"]]
2377 + : [],
2378 + ),
2379 + [
2380 + `/${PI_T3_SERVICE_TIER_COMMAND} priority`,
2381 + `/${PI_T3_SERVICE_TIER_COMMAND} priority`,
2382 + `/${PI_T3_SERVICE_TIER_COMMAND} ultrafast`,
2383 + `/${PI_T3_SERVICE_TIER_COMMAND}`,
2384 + ],
2385 + );
2386 + }).pipe(Effect.scoped, Effect.provide(layerTest)),
2387 + );
2388 +
2389 it.effect.each([false, true, "previous"] as const)(
2390 "resets applied thinking when returning to Pi default (failed first restore: %s)",
2391 (failFirstRestore) =>
packages/provider-pi/src/server/adapter.ts
+25 -1
@@ -71,7 +71,8 @@ import {
71 materializePiT3McpExtension,
72 resolvePiLaunchArgs,
73 } from "./mcpInjection.ts";
74 -import { PI_FILE_CHANGE_TOOLS } from "./mcpExtensionSource.ts";
74 +import { PI_FILE_CHANGE_TOOLS, PI_T3_SERVICE_TIER_COMMAND } from "./mcpExtensionSource.ts";
75 +import { PI_STANDARD_SERVICE_TIER } from "./modelCapabilities.ts";
76
77 export const PI_PROVIDER = ProviderDriverKind.make("pi");
78 const PI_DRIVER_KIND = PI_PROVIDER;
@@ -575,6 +576,9 @@ export const makePiAdapterV2 = Effect.fn("makePiAdapterV2")(function* (
576 );
577 let appliedModel: string | null = null;
578 let appliedThinking: string | null = null;
579 + // Pi reloads extensions (resetting the tier) on paths the adapter cannot see, so a chosen tier
580 + // is resent every turn; Standard sends a clear only when an earlier tier may still be set.
581 + let serviceTierMayBeSet = false;
582 let appliedSessionName: string | null = null;
583 const outOfTurnExtensionErrors: Array<PiRpcRecord> = [];
584 let lastKnownLeaf: string | null = null;
@@ -2200,6 +2204,8 @@ export const makePiAdapterV2 = Effect.fn("makePiAdapterV2")(function* (
2204 }
2205 return;
2206 }
2207 + // Only id-less prompts own ACK slots; a late reply to an id-bearing request must not take one.
2208 + if (command === "prompt" && typeof event["id"] === "string") return;
2209 const pendingPrompt = command === "prompt" ? pendingPromptResponses.shift() : undefined;
2210 const responseTurn =
2211 pendingPrompt?.providerTurnId === turn?.providerTurn.id ? turn : null;
@@ -2598,6 +2604,24 @@ export const makePiAdapterV2 = Effect.fn("makePiAdapterV2")(function* (
2604 yield* request({ type: "set_thinking_level", level: thinking });
2605 appliedThinking = thinking;
2606 }
2607 + const requestedTier = getModelSelectionStringOptionValue(modelSelection, "serviceTier");
2608 + const serviceTier =
2609 + requestedTier === undefined || requestedTier === PI_STANDARD_SERVICE_TIER
2610 + ? null
2611 + : requestedTier;
2612 + if (serviceTier !== null || serviceTierMayBeSet) {
2613 + serviceTierMayBeSet = true;
2614 + const result = yield* request({
2615 + type: "prompt",
2616 + message: `/${PI_T3_SERVICE_TIER_COMMAND} ${serviceTier ?? ""}`.trimEnd(),
2617 + });
2618 + // Pi before 0.99 acknowledges a handled extension command without a disposition.
2619 + const disposition = recordString(result, "disposition");
2620 + if (disposition !== undefined && disposition !== "handled") {
2621 + return yield* protocolError("T3 Code's Pi extension did not apply the service tier");
2622 + }
2623 + serviceTierMayBeSet = serviceTier !== null;
2624 + }
2625 });
2626
2627 const resolvePromptPayload = Effect.fnUntraced(function* (
packages/provider-pi/src/server/mcpBridge.testkit.ts
+1
@@ -125,6 +125,7 @@ export async function loadMcpBridge(
125 },
126 pi: {
127 on: (name: string, handler: AgentStartHook) => handlers.set(name, handler),
128 + registerCommand: () => {},
129 registerTool: (tool: RegisteredTool) => {
130 if (options.allowsTool && !options.allowsTool(tool.name)) return;
131 const index = tools.findIndex((current) => current.name === tool.name);
packages/provider-pi/src/server/mcpExtensionSource.test.ts
+40 -5
@@ -6,14 +6,16 @@ import * as NodeModule from "node:module";
6 import * as NodeVM from "node:vm";
7 import { assert, describe, it } from "@effect/vitest";
8
9 -import { PI_T3_MCP_EXTENSION_SOURCE } from "./mcpExtensionSource.ts";
9 +import { PI_T3_MCP_EXTENSION_SOURCE, PI_T3_SERVICE_TIER_COMMAND } from "./mcpExtensionSource.ts";
10 import { loadMcpBridge } from "./mcpBridge.testkit.ts";
11
12 type RequestHook = (
13 event: { payload: unknown },
14 - ctx: { model: { provider: string } },
14 + ctx: { model: { provider: string; serviceTiers?: ReadonlyArray<{ id: string }> } },
15 ) => Record<string, unknown> | undefined;
16
17 +type CommandHandler = (args: string) => Promise<void>;
18 +
19 interface SkillCommand {
20 readonly name: string;
21 readonly source: string;
@@ -43,6 +45,7 @@ async function loadHooks(
45 stripFrontmatter: (content: string) => content,
46 pi: {
47 on: (name: string, handler: unknown) => handlers.set(name, handler),
48 + registerCommand: () => {},
49 getCommands: () => commands,
50 },
51 });
@@ -367,6 +370,7 @@ describe("Pi tool discovery permissions", () => {
370 on: (name: string, handler: ToolCallHook) => {
371 if (name === "tool_call") toolCall = handler;
372 },
373 + registerCommand: () => {},
374 getAllTools: () => [{ name: "tool_search", sourceInfo: { path: searchPath } }],
375 },
376 });
@@ -398,8 +402,9 @@ describe("Pi tool discovery permissions", () => {
402 });
403 });
404
401 -async function loadRequestHook(): Promise<RequestHook> {
405 +async function loadRequestHook() {
406 const handlers = new Map<string, RequestHook>();
407 + const commands = new Map<string, CommandHandler>();
408 const source = NodeModule.stripTypeScriptTypes(
409 PI_T3_MCP_EXTENSION_SOURCE.replace(/^import .*;$/gm, "").replace(
410 "export default async function",
@@ -408,13 +413,43 @@ async function loadRequestHook(): Promise<RequestHook> {
413 );
414 await NodeVM.runInNewContext(`${source}\nt3McpExtension(pi)`, {
415 process: { env: {} },
411 - pi: { on: (name: string, handler: RequestHook) => handlers.set(name, handler) },
416 + pi: {
417 + on: (name: string, handler: RequestHook) => handlers.set(name, handler),
418 + registerCommand: (name: string, options: { handler: CommandHandler }) =>
419 + commands.set(name, options.handler),
420 + },
421 });
422 const hook = handlers.get("before_provider_request");
423 + const setServiceTier = commands.get(PI_T3_SERVICE_TIER_COMMAND);
424 assert.isDefined(hook);
415 - return hook!;
425 + assert.isDefined(setServiceTier);
426 + return Object.assign(hook!, { setServiceTier: setServiceTier! });
427 }
428
429 +describe("Pi service tier", () => {
430 + it("sends only a tier the active model advertises, until Standard clears it", async () => {
431 + const hook = await loadRequestHook();
432 + const sol = {
433 + provider: "columbia-codex",
434 + serviceTiers: [{ id: "priority" }, { id: "ultrafast" }],
435 + };
436 + const payload = { model: "gpt-6.1-sol", reasoning: { effort: "high" } };
437 + assert.isUndefined(hook({ payload }, { model: sol }));
438 +
439 + await hook.setServiceTier("priority");
440 + assert.deepEqual(hook({ payload }, { model: sol }), { ...payload, service_tier: "priority" });
441 +
442 + await hook.setServiceTier("ultrafast");
443 + assert.equal(hook({ payload }, { model: sol })?.service_tier, "ultrafast");
444 + const fastOnly = { provider: "columbia-codex", serviceTiers: [{ id: "priority" }] };
445 + assert.isUndefined(hook({ payload }, { model: fastOnly }));
446 + assert.isUndefined(hook({ payload }, { model: { provider: "anthropic" } }));
447 +
448 + await hook.setServiceTier("");
449 + assert.isUndefined(hook({ payload }, { model: sol }));
450 + });
451 +});
452 +
453 describe("Pi upstream output-budget workaround", () => {
454 it.each(["max_tokens", "max_completion_tokens"])(
455 "caps %s without changing the conversation or tools",
packages/provider-pi/src/server/mcpExtensionSource.ts
+33 -11
@@ -7,6 +7,7 @@ export const T3_MCP_URL_ENV = "T3_MCP_URL";
7 export const T3_MCP_BEARER_ENV = "T3_MCP_BEARER_TOKEN";
8 export const T3_PI_RUNTIME_MODE_ENV = "T3_PI_RUNTIME_MODE";
9 export const T3_PI_MCP_EXTENSION_PATH_ENV = "T3_PI_MCP_EXTENSION_PATH";
10 +export const PI_T3_SERVICE_TIER_COMMAND = "t3-service-tier";
11
12 // These tools trigger file-change approvals, bypassed by Auto-accept edits.
13 export const PI_FILE_CHANGE_TOOLS = ["edit", "write"] as const;
@@ -22,6 +23,7 @@ const TOKEN_ENV = ${JSON.stringify(T3_MCP_BEARER_ENV)};
23 const RUNTIME_MODE_ENV = ${JSON.stringify(T3_PI_RUNTIME_MODE_ENV)};
24 const EXTENSION_PATH_ENV = ${JSON.stringify(T3_PI_MCP_EXTENSION_PATH_ENV)};
25 const ORCHESTRATION_INSTRUCTIONS = ${JSON.stringify(T3_CODE_ORCHESTRATION_INSTRUCTIONS.trim())};
26 +const SERVICE_TIER_COMMAND = ${JSON.stringify(PI_T3_SERVICE_TIER_COMMAND)};
27 const PROTOCOL = "2025-06-18";
28 const READ_ONLY_TOOLS = new Set(["read", "grep", "find", "ls"]);
29 const FILE_CHANGE_TOOLS = new Set(${JSON.stringify(PI_FILE_CHANGE_TOOLS)});
@@ -287,24 +289,44 @@ export default async function t3McpExtension(pi: ExtensionAPI) {
289 };
290 });
291
290 - // Workaround for an upstream Pi context-budgeting bug: pi-ai reuses the
291 - // previous response's usage even when a fork's instructions/tools differ,
292 - // then reserves almost all remaining context for output. OpenRouter can
293 - // reject even a short conversation. Remove this cap when Pi accounts for
294 - // the current request prefix reliably (api/simple-options + utils/estimate).
292 + // Pi has no service tier setting, so T3 sets one through this command before each turn.
293 + // Only a tier the active model advertises is sent; Standard sends none.
294 + let serviceTier: string | undefined;
295 + pi.registerCommand(SERVICE_TIER_COMMAND, {
296 + description: "Set the service tier T3 Code sends with provider requests",
297 + handler: async (args) => {
298 + serviceTier = args.trim() || undefined;
299 + },
300 + });
301 +
302 pi.on("before_provider_request", (event, ctx) => {
296 - if (ctx.model?.provider !== "openrouter") return;
303 const payload = event.payload;
304 if (typeof payload !== "object" || payload === null || Array.isArray(payload)) return;
305 const replacement = { ...payload } as Record<string, unknown>;
306 let changed = false;
301 - for (const key of ["max_tokens", "max_completion_tokens"]) {
302 - const limit = replacement[key];
303 - if (typeof limit === "number" && Number.isFinite(limit) && limit > 32_768) {
304 - replacement[key] = 32_768;
305 - changed = true;
307 + // Workaround for an upstream Pi context-budgeting bug: pi-ai reuses the
308 + // previous response's usage even when a fork's instructions/tools differ,
309 + // then reserves almost all remaining context for output. OpenRouter can
310 + // reject even a short conversation. Remove this cap when Pi accounts for
311 + // the current request prefix reliably (api/simple-options + utils/estimate).
312 + if (ctx.model?.provider === "openrouter") {
313 + for (const key of ["max_tokens", "max_completion_tokens"]) {
314 + const limit = replacement[key];
315 + if (typeof limit === "number" && Number.isFinite(limit) && limit > 32_768) {
316 + replacement[key] = 32_768;
317 + changed = true;
318 + }
319 }
320 }
321 + const tiers = (ctx.model as { serviceTiers?: unknown } | undefined)?.serviceTiers;
322 + if (
323 + serviceTier !== undefined &&
324 + Array.isArray(tiers) &&
325 + tiers.some((tier) => tier?.id === serviceTier)
326 + ) {
327 + replacement.service_tier = serviceTier;
328 + changed = true;
329 + }
330 if (changed) return replacement;
331 });
332
packages/provider-pi/src/server/modelCapabilities.ts renamed
+62 -17
@@ -1,4 +1,8 @@
1 -import { type ModelCapabilities, type ProviderOptionChoice } from "@t3tools/contracts";
1 +import {
2 + type ModelCapabilities,
3 + type ProviderOptionChoice,
4 + type ProviderOptionDescriptor,
5 +} from "@t3tools/contracts";
6 import { createModelCapabilities } from "@t3tools/shared/model";
7 import * as Predicate from "effect/Predicate";
8
@@ -21,27 +25,68 @@ export const EMPTY_PI_MODEL_CAPABILITIES: ModelCapabilities = createModelCapabil
25 optionDescriptors: [],
26 });
27
24 -export function thinkingCapabilitiesForPiModel(
28 +/** Selecting Standard sends no `service_tier`, leaving the endpoint's default. */
29 +export const PI_STANDARD_SERVICE_TIER = "default";
30 +
31 +export function piModelCapabilities(
32 model: unknown,
33 defaultThinkingLevel: unknown,
34 ): ModelCapabilities {
35 + const optionDescriptors = [
36 + thinkingDescriptor(model, defaultThinkingLevel),
37 + serviceTierDescriptor(model),
38 + ].filter(Predicate.isNotUndefined);
39 + return optionDescriptors.length === 0
40 + ? EMPTY_PI_MODEL_CAPABILITIES
41 + : createModelCapabilities({ optionDescriptors });
42 +}
43 +
44 +function thinkingDescriptor(
45 + model: unknown,
46 + defaultThinkingLevel: unknown,
47 +): ProviderOptionDescriptor | undefined {
48 const levels = supportedPiThinkingLevelsFromModel(model);
29 - if (levels.length === 0) return EMPTY_PI_MODEL_CAPABILITIES;
49 + if (levels.length === 0) return undefined;
50 const defaultLevel = clampPiThinkingLevel(defaultThinkingLevel, levels);
31 - return createModelCapabilities({
32 - optionDescriptors: [
33 - {
34 - id: "thinking",
35 - label: "Thinking",
36 - type: "select",
37 - options: levels.map((level): ProviderOptionChoice => ({
38 - id: level,
39 - label: PI_THINKING_LEVEL_LABELS[level],
40 - ...(level === defaultLevel ? { isDefault: true } : {}),
41 - })),
42 - },
43 - ],
44 - });
51 + return {
52 + id: "thinking",
53 + label: "Thinking",
54 + type: "select",
55 + options: levels.map((level): ProviderOptionChoice => ({
56 + id: level,
57 + label: PI_THINKING_LEVEL_LABELS[level],
58 + ...(level === defaultLevel ? { isDefault: true } : {}),
59 + })),
60 + };
61 +}
62 +
63 +// Pi has no service tier concept; provider extensions advertise `serviceTiers` on a model,
64 +// and T3's injected extension sends the chosen id as the request's `service_tier`.
65 +function serviceTierDescriptor(model: unknown): ProviderOptionDescriptor | undefined {
66 + const tiers = recordField(model, "serviceTiers");
67 + if (!Array.isArray(tiers)) return undefined;
68 + const seen = new Set<string>([PI_STANDARD_SERVICE_TIER]);
69 + const options: Array<ProviderOptionChoice> = [];
70 + for (const tier of tiers) {
71 + const id = recordField(tier, "id");
72 + if (typeof id !== "string" || id.trim() === "" || seen.has(id)) continue;
73 + seen.add(id);
74 + const name = recordField(tier, "name");
75 + const description = recordField(tier, "description");
76 + options.push({
77 + id,
78 + label: typeof name === "string" && name.trim() !== "" ? name : id,
79 + ...(typeof description === "string" && description.trim() !== "" ? { description } : {}),
80 + });
81 + }
82 + if (options.length === 0) return undefined;
83 + return {
84 + id: "serviceTier",
85 + label: "Service Tier",
86 + type: "select",
87 + options: [{ id: PI_STANDARD_SERVICE_TIER, label: "Standard", isDefault: true }, ...options],
88 + currentValue: PI_STANDARD_SERVICE_TIER,
89 + };
90 }
91
92 function clampPiThinkingLevel(
packages/provider-pi/src/server/status.test.ts
+43
@@ -276,6 +276,49 @@ describe("PiProvider", () => {
276 }).pipe(Effect.provide(NodeServices.layer)),
277 );
278
279 + it.effect("advertises a model's service tiers beside its thinking levels", () =>
280 + Effect.gen(function* () {
281 + const spawner = yield* piDiscoverySpawner({
282 + models: [
283 + {
284 + provider: "columbia-codex",
285 + id: "gpt-6.1-sol",
286 + reasoning: true,
287 + serviceTiers: [
288 + { id: "priority", name: "Fast", description: "2x speed, increased usage" },
289 + { id: "ultrafast", name: "Ultrafast" },
290 + ],
291 + },
292 + { provider: "anthropic", id: "claude-sonnet-4", reasoning: true },
293 + ],
294 + });
295 + const snapshot = yield* checkPiProviderStatus(settings).pipe(
296 + Effect.provideService(ChildProcessSpawner.ChildProcessSpawner, spawner),
297 + );
298 + const descriptors = (slug: string) =>
299 + snapshot.models.find((model) => model.slug === slug)?.capabilities?.optionDescriptors ?? [];
300 + assert.deepEqual(
301 + descriptors("columbia-codex/gpt-6.1-sol").map((descriptor) => descriptor.id),
302 + ["thinking", "serviceTier"],
303 + );
304 + assert.deepEqual(descriptors("columbia-codex/gpt-6.1-sol")[1], {
305 + id: "serviceTier",
306 + label: "Service Tier",
307 + type: "select",
308 + options: [
309 + { id: "default", label: "Standard", isDefault: true },
310 + { id: "priority", label: "Fast", description: "2x speed, increased usage" },
311 + { id: "ultrafast", label: "Ultrafast" },
312 + ],
313 + currentValue: "default",
314 + });
315 + assert.deepEqual(
316 + descriptors("anthropic/claude-sonnet-4").map((descriptor) => descriptor.id),
317 + ["thinking"],
318 + );
319 + }).pipe(Effect.provide(NodeServices.layer)),
320 + );
321 +
322 // regression: #18498 a malformed model inventory must not restore stale Pi models
323 it.effect.each([
324 ["missing models", {}],
packages/provider-pi/src/server/status.ts
+2 -5
@@ -40,10 +40,7 @@ import {
40 enrichProviderSnapshotWithVersionAdvisory,
41 type ProviderMaintenanceCapabilities,
42 } from "@t3tools/provider-core/server/maintenanceResolver";
43 -import {
44 - EMPTY_PI_MODEL_CAPABILITIES,
45 - thinkingCapabilitiesForPiModel,
46 -} from "./thinkingCapabilities.ts";
43 +import { EMPTY_PI_MODEL_CAPABILITIES, piModelCapabilities } from "./modelCapabilities.ts";
44 import {
45 parsePiDiscoveredCommands,
46 withPiBuiltinSlashCommands,
@@ -127,7 +124,7 @@ function parseDiscoveredModels(
124 name: recordString(model, "name") ?? slug,
125 subProvider: provider,
126 isCustom: false,
130 - capabilities: thinkingCapabilitiesForPiModel(
127 + capabilities: piModelCapabilities(
128 model,
129 provider === activeProvider && id === activeModelId
130 ? recordString(state, "thinkingLevel")