feat: the model picker marks LiteLLM gateway models, and Pi leaves the LiteLLM launcher

apps/mobile/src/features/threads/ThreadSettingsRows.shared.tsx

@@ -80,6 +80,11 @@ export function ModelRowContent(
                 <Text className="text-3xs font-t3-bold text-foreground-muted">Legacy</Text>
               </View>
             ) : null}
+            {props.option.isGateway ? (
+              <View className="rounded-md bg-subtle px-1.5 py-0.5">
+                <Text className="text-3xs font-t3-bold text-foreground-muted">LiteLLM</Text>
+              </View>
+            ) : null}
             {props.option.isHidden ? (
               <Text className="text-xs text-foreground-muted">Hidden</Text>
             ) : null}

apps/mobile/src/lib/modelOptions.ts

@@ -40,6 +40,8 @@ export type ModelOption = {
   };
   readonly isDefault: boolean;
   readonly isLegacy: boolean;
+  /** Served by a LiteLLM gateway rather than the harness's own catalog. */
+  readonly isGateway?: boolean;
   readonly isUnavailable?: boolean;
   readonly isHidden?: boolean;
   readonly canHide?: boolean;
@@ -381,6 +383,7 @@ export function buildModelOptions(
         canHide: !model.isCustom,
         isDefault: model.isDefault === true,
         isLegacy: model.isLegacy === true,
+        ...(model.gateway ? { isGateway: true } : {}),
         ...(updateRequired ? { providerUpdateRequired: updateRequired } : {}),
         capabilities: model.capabilities,
         selection: normalizeSelectionOptions(
@@ -444,6 +447,7 @@ export function buildModelOptions(
         providerBadge: providerBadgeFor(fallbackModelSelection.instanceId),
         isDefault: false,
         isLegacy: model?.isLegacy === true,
+        ...(model?.gateway ? { isGateway: true } : {}),
         ...(isModelSelectionUnavailable(config, fallbackModelSelection, projectId) ||
         (provider !== undefined && !isProviderEnabledForProject(config, provider, projectId)) ||
         provider?.updateRequiredModels?.some((gated) => gated.slug === fallbackModelSelection.model)

apps/server/src/orchestration-v2/ProviderAuthRetry.ts

@@ -17,7 +17,7 @@ import * as ThreadManagementService from "./ThreadManagementService.ts";
 const RETRY_MESSAGE_PREFIX = "message:auth-retry:";
 
 const AUTH_REJECTION_CODES = new Set(["api_error_401", "api_error_403"]);
-// Codex, OpenCode and Pi only pass the gateway's text through. LiteLLM answers an
+// Codex and OpenCode only pass the gateway's text through. LiteLLM answers an
 // expired or revoked key with "Authentication Error, ..." and HTTP 401.
 // A bare 401 or "unauthorized" is too common in tool output to count.
 const AUTH_REJECTION_TEXT =

apps/server/src/provider/builtInDrivers.ts

@@ -21,7 +21,7 @@ import {
   type LiteLLMLauncherEnv,
 } from "@t3tools/provider-core/server/litellm";
 
-import { LITELLM_CLAUDE, LITELLM_CODEX, LITELLM_OPENCODE, LITELLM_PI } from "./liteLLMLaunchers.ts";
+import { LITELLM_CLAUDE, LITELLM_CODEX, LITELLM_OPENCODE } from "./liteLLMLaunchers.ts";
 
 export type BuiltInDriversEnv =
   | AcpRegistryDriverEnv
@@ -46,7 +46,7 @@ export const BUILT_IN_DRIVERS: ReadonlyArray<AnyProviderDriver<BuiltInDriversEnv
   withLiteLLMLauncher(OpenCodeDriver, LITELLM_OPENCODE),
   AntigravityDriver,
   KiroDriver,
-  withLiteLLMLauncher(PiDriver, LITELLM_PI),
+  PiDriver,
   MuseDriver,
   AcpRegistryDriver,
 ];

apps/server/src/provider/liteLLMLaunchers.ts

@@ -1,7 +1,6 @@
 import type { ClaudeSettings, CodexSettings } from "@t3tools/contracts";
 import type { LiteLLMLauncherSpec } from "@t3tools/provider-core/server/litellm";
 import type { OpenCodeSettings } from "@t3tools/provider-opencode/settings";
-import type { PiSettings } from "@t3tools/provider-pi/settings";
 import { tokenizeCliArgs } from "@t3tools/shared/cliArgs";
 import { codexModelFamily } from "@t3tools/shared/model";
 import * as Effect from "effect/Effect";
@@ -41,16 +40,17 @@ export const LITELLM_CLAUDE: LiteLLMLauncherSpec<ClaudeSettings> = {
   gatewayModelSlug: claudeGatewayModelSlug,
 };
 
+const OPENCODE_GATEWAY_PREFIX = "litellm/";
+
 export const LITELLM_OPENCODE: LiteLLMLauncherSpec<OpenCodeSettings> = {
   harness: "opencode",
   configure: (config, launch) =>
     config.serverUrl
       ? Effect.fail("Clear Server URL: LiteLLM can only route an OpenCode server T3 starts.")
       : Effect.succeed({ ...config, binaryPath: launch.binaryPath }),
-};
-
-export const LITELLM_PI: LiteLLMLauncherSpec<PiSettings> = {
-  harness: "pi",
-  configure: (config, launch) => Effect.succeed({ ...config, binaryPath: launch.binaryPath }),
-  homeEnvironment: (homeDir) => ({ PI_CODING_AGENT_DIR: homeDir }),
+  // `lite opencode` adds the gateway as the `litellm` provider; other providers stay direct.
+  listedModelGatewayId: (slug) =>
+    slug.startsWith(OPENCODE_GATEWAY_PREFIX)
+      ? slug.slice(OPENCODE_GATEWAY_PREFIX.length)
+      : undefined,
 };

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

@@ -29,6 +29,8 @@ export const ModelListRow = memo(function ModelListRow(props: {
   /** Marks this instance apart from another of the same provider. */
   showInstanceBadge: boolean;
   showNewBadge?: boolean;
+  /** Served by a LiteLLM gateway rather than the harness's own catalog. */
+  showGatewayBadge?: boolean;
   unavailable?: boolean;
   jumpLabel?: string | null;
   disabledReason?: string | null;
@@ -93,6 +95,11 @@ export const ModelListRow = memo(function ModelListRow(props: {
             New
           </span>
         ) : null}
+        {props.showGatewayBadge ? (
+          <Badge variant="outline" size="sm" aria-label="Served by the LiteLLM gateway">
+            LiteLLM
+          </Badge>
+        ) : null}
         {props.unavailable ? (
           <Badge variant="outline" size="sm">
             Unavailable

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

@@ -79,6 +79,7 @@ type ModelPickerItem = {
   acpRegistryIconUrl?: string | undefined;
   continuationGroupKey?: string | undefined;
   isLegacy?: boolean | undefined;
+  gateway?: "litellm" | undefined;
   isUnavailable?: boolean | undefined;
 };
 
@@ -567,6 +568,7 @@ export const ModelPickerContent = memo(function ModelPickerContent(props: {
           ...(model.subProvider ? { subProvider: model.subProvider } : {}),
           ...(model.badge ? { badge: model.badge } : {}),
           ...(model.isLegacy ? { isLegacy: true } : {}),
+          ...(model.gateway ? { gateway: model.gateway } : {}),
           ...(model.isUnavailable ? { isUnavailable: true } : {}),
           instanceId,
           driverKind: entry.driverKind,
@@ -1056,6 +1058,7 @@ export const ModelPickerContent = memo(function ModelPickerContent(props: {
                     favoriteStarOnHover={selectedInstanceId === "favorites" && !isSearching}
                     showInstanceBadge={badgedInstanceIds.has(model.instanceId)}
                     showNewBadge={model.badge === "new"}
+                    showGatewayBadge={model.gateway === "litellm"}
                     unavailable={model.isUnavailable === true}
                     jumpLabel={modelJumpLabelByKey.get(modelKey) ?? null}
                     disabledReason={disabledReason}

apps/web/src/components/chat/providerIconUtils.ts

@@ -10,6 +10,7 @@ export type ModelEsque = {
   isDefault?: boolean | undefined;
   badge?: "new" | undefined;
   isLegacy?: boolean | undefined;
+  gateway?: "litellm" | undefined;
   isUnavailable?: boolean | undefined;
 };
 

apps/web/src/components/settings/AddProviderInstanceDialog.logic.ts

@@ -27,7 +27,6 @@ export const LITELLM_LAUNCHER_DRIVERS: ReadonlySet<string> = new Set([
   "claudeAgent",
   "codex",
   "opencode",
-  "pi",
 ]);
 
 /** The suggested label follows the launcher until the user types their own. */

apps/web/src/components/settings/ProviderInstanceCard.tsx

@@ -1287,7 +1287,7 @@ export function ProviderInstanceCard({
             </div>
           }
         />
-        {LITELLM_LAUNCHER_DRIVERS.has(instance.driver) ? (
+        {LITELLM_LAUNCHER_DRIVERS.has(instance.driver) || instance.launcher === "litellm" ? (
           <SettingsRow
             title={
               <label htmlFor={`provider-instance-${instanceId}-litellm`}>

apps/web/src/components/settings/ProviderModelsSection.tsx

@@ -532,6 +532,9 @@ function ProviderModelRow({
         {model.isCustom ? (
           <span className="shrink-0 text-2xs text-muted-foreground/70">custom</span>
         ) : null}
+        {model.gateway === "litellm" ? (
+          <span className="shrink-0 text-2xs text-muted-foreground/70">LiteLLM</span>
+        ) : null}
       </span>
       <span className="text-2xs text-muted-foreground/70">
         {capLabels.length > 0 ? (

apps/web/src/modelSelection.ts

@@ -70,6 +70,7 @@ export interface AppModelOption {
   isCustom: boolean;
   isDefault?: boolean;
   isLegacy?: boolean;
+  gateway?: "litellm";
   isUnavailable?: boolean;
 }
 
@@ -80,7 +81,7 @@ const LITELLM_REQUEST_TAG_DRIVERS: ReadonlySet<string> = new Set([
   "opencode",
 ]);
 
-/** LiteLLM tags reach the gateway through Claude, Codex and OpenCode; Pi has no tag channel. */
+/** LiteLLM tags reach the gateway through Claude, Codex and OpenCode. */
 export function supportsLiteLLMRequestTags(
   settings: Pick<UnifiedSettings, "providerInstances">,
   instanceId: ProviderInstanceId,
@@ -150,6 +151,7 @@ function toAppModelOption(
   if (model.badge) option.badge = model.badge;
   if (model.isDefault) option.isDefault = true;
   if (model.isLegacy) option.isLegacy = true;
+  if (model.gateway) option.gateway = model.gateway;
   return option;
 }
 

docs/user/providers-litellm.md

@@ -1,6 +1,6 @@
 # LiteLLM
 
-T3 Code can run Claude, Codex, OpenCode, and Pi through a [LiteLLM](https://docs.litellm.ai)
+T3 Code can run Claude, Codex, and OpenCode through a [LiteLLM](https://docs.litellm.ai)
 gateway with the LiteLLM CLI (`lite`). Each harness keeps working as usual, but it signs in
 with your gateway's SSO and uses the gateway's models. This lets you, for example, use a
 GPT model on the Claude harness.
@@ -17,7 +17,7 @@ GPT model on the Claude harness.
 2. Set `LITELLM_PROXY_URL` to your gateway URL in your shell profile, for example
    `export LITELLM_PROXY_URL=https://litellm.example.com`. The desktop app reads it from your
    login shell. You can also set it in a provider instance's **Environment variables**.
-3. In **Settings > Providers**, choose **Add provider**, pick Claude, Codex, OpenCode, or Pi,
+3. In **Settings > Providers**, choose **Add provider**, pick Claude, Codex, or OpenCode,
    and turn on **Launch through LiteLLM**. The suggested label is, for example,
    **LiteLLM Claude**, and that is the name the model picker shows. To route an existing
    provider instead, such as the built-in Codex or Claude, open it and turn on
@@ -42,8 +42,12 @@ full URL and paste it in T3 Code when it asks.
 ## Models
 
 - **LiteLLM Claude** lists every model the gateway offers.
-- **LiteLLM Codex**, **LiteLLM OpenCode**, and **LiteLLM Pi** list the models that `lite`
-  syncs into those harnesses.
+- **LiteLLM Codex** and **LiteLLM OpenCode** list the models that `lite` syncs into those
+  harnesses.
+
+In the model picker and in **Settings > Providers**, a **LiteLLM** label marks each model the
+gateway serves. A model without the label comes from the harness's own list. LiteLLM OpenCode
+sends those models to their usual provider, not to the gateway.
 
 T3 Code checks the gateway's model list at most once a minute, and **Refresh** in
 **Settings > Providers** checks it now. If the gateway removes the model a thread uses, the
@@ -71,13 +75,11 @@ commit messages, and other text generation do not carry tags.
 - Threads that a thread starts, such as delegated tasks, forks, and threads an agent creates,
   copy its tags. A scheduled task keeps the tags you set on it.
 
-LiteLLM Pi does not send request tags yet.
-
 ## Limits
 
-- LiteLLM Codex and LiteLLM Pi use their own config directory, so they do not read your
-  usual Codex or Pi config, skills, or extensions. A Codex provider you switch to LiteLLM
-  might not resume threads it started before the switch.
+- LiteLLM Codex uses its own config directory, so it does not read your usual Codex config or
+  skills. A Codex provider you switch to LiteLLM might not resume threads it started before
+  the switch.
 - When the selected text generation model is not on the gateway, commit messages, titles,
   and branch names use the first gateway model.
 - LiteLLM Codex does not accept `-c` or `--config` in **Launch arguments**, because Codex

packages/contracts/src/server.ts

@@ -83,6 +83,8 @@ export const ServerProviderModel = Schema.Struct({
   isCustom: Schema.Boolean,
   isDefault: Schema.optional(Schema.Boolean),
   isLegacy: Schema.optional(Schema.Boolean),
+  /** Set when a LiteLLM gateway serves the model, as opposed to the harness's own catalog. */
+  gateway: Schema.optional(Schema.Literal("litellm")),
   capabilities: Schema.NullOr(ModelCapabilities),
 });
 export type ServerProviderModel = typeof ServerProviderModel.Type;

packages/provider-core/src/server/litellm.test.ts

@@ -35,6 +35,7 @@ import {
   liteLLMSignedOutMessage,
   parseLiteLLMModelIds,
   type LiteLLMHarness,
+  type LiteLLMStatus,
   withLiteLLMLauncher,
 } from "./litellm.ts";
 import type {
@@ -92,20 +93,44 @@ describe("applyLiteLLMStatus", () => {
       (id) => id.replace(/^bedrock\//, ""),
     );
     assert.deepEqual(
-      result.models.map((model) => [model.slug, model.name, model.isCustom]),
+      result.models.map((model) => [model.slug, model.name, model.isCustom, model.gateway]),
       [
-        ["gpt-5", "gpt-5", false],
-        ["claude-sonnet-4-5", "Claude Sonnet 4.5", false],
-        ["bedrock/sonnet-4.5", "bedrock/sonnet-4.5", false],
-        ["claude sonnet 4.5", "claude sonnet 4.5", false],
-        ["my-alias", "My alias", true],
+        ["gpt-5", "gpt-5", false, "litellm"],
+        ["claude-sonnet-4-5", "Claude Sonnet 4.5", false, "litellm"],
+        ["bedrock/sonnet-4.5", "bedrock/sonnet-4.5", false, "litellm"],
+        ["claude sonnet 4.5", "claude sonnet 4.5", false, "litellm"],
+        ["my-alias", "My alias", true, undefined],
       ],
     );
     assert.strictEqual(result.models[0]?.capabilities, capabilities);
     assert.strictEqual(result.models[2]?.capabilities, sonnetCapabilities);
     assert.strictEqual(result.models[3]?.capabilities, sonnetCapabilities);
   });
 
+  it("marks only the harness-listed models the gateway serves", () => {
+    const listed = (slug: string) => ({ slug, name: slug, isCustom: false, capabilities: null });
+    const harnessSnapshot = snapshot({
+      models: [listed("litellm/gpt-5"), listed("openai/gpt-5"), listed("litellm/retired")],
+    });
+    const gatewayId = (slug: string) =>
+      slug.startsWith("litellm/") ? slug.slice("litellm/".length) : undefined;
+    const marks = (status: LiteLLMStatus) =>
+      applyLiteLLMStatus(harnessSnapshot, status, undefined, undefined, gatewayId).models.map(
+        (model) => [model.slug, model.gateway],
+      );
+    assert.deepEqual(marks({ auth: { status: "authenticated" }, models: ["gpt-5"] }), [
+      ["litellm/gpt-5", "litellm"],
+      ["openai/gpt-5", undefined],
+      ["litellm/retired", undefined],
+    ]);
+    // Without the gateway's list, no model can claim to be served by it.
+    assert.deepEqual(marks({ auth: { status: "unauthenticated", message: "Signed out." } }), [
+      ["litellm/gpt-5", undefined],
+      ["openai/gpt-5", undefined],
+      ["litellm/retired", undefined],
+    ]);
+  });
+
   it("reports the gateway sign-in instead of the harness's own account", () => {
     const signedOut = applyLiteLLMStatus(
       snapshot(),
@@ -326,7 +351,6 @@ const createLiteLLMInstance = (
     harness: options.harness ?? "claude",
     configure: (config, launch) =>
       Effect.succeed({ ...config, binaryPath: launch.binaryPath, homePath: launch.homeDir }),
-    homeEnvironment: (homeDir) => ({ FAKE_HOME: homeDir }),
     gatewayModelCapabilities: capabilities,
     servesModel: (model, requested) => model.slug === `my-${requested}`,
   });
@@ -387,10 +411,6 @@ describe.skipIf(windowsHost)("withLiteLLMLauncher", () => {
         shim,
         NodePath.join(fixture.dir, "state", "litellm", instanceId, "bin", "claude"),
       );
-      assert.deepEqual(
-        input.environment.find((variable) => variable.name === "FAKE_HOME")?.value,
-        input.config.homePath,
-      );
 
       const run = (args: ReadonlyArray<string>) =>
         spawnAndCollect(shim, ChildProcess.make(shim, [...args], { env: fixture.environment }));
@@ -603,8 +623,6 @@ describe.skipIf(windowsHost)("withLiteLLMLauncher", () => {
         "tool:company,project:billing",
       ]);
       assert.deepEqual(yield* tagsWith({ mode: "off", custom: "" }), ["", "project:billing"]);
-      const pi = yield* createLiteLLMInstance(fixture, [], { harness: "pi" });
-      assert.isUndefined(pi.orchestrationAdapter.launchTags);
     }).pipe(Effect.scoped, Effect.provide(testServices)),
   );