thursday-agent 0.2.0 → 0.3.0

package/features/ai/model.schema.ts

@@ -1,4 +1,5 @@
 import { z } from "zod";
+import { BOT_RUN } from "@/config";
 import {
   REALTIME_PROVIDERS,
   type SpeachModelProviderId,
@@ -8,6 +9,7 @@ import {
 
 export const textModelProviderSchema = z.enum([
   "openai",
+  "chatgpt",
   "anthropic",
   "google",
   "xai",
@@ -42,7 +44,46 @@ export const MODEL_TIERS = ["small", "mid", "large"] as const;
 export type ModelTier = (typeof MODEL_TIERS)[number];
 
 /** One model worth offering, as the picker reads it. */
-export type SuggestModel = { id: string; label: string; tier: ModelTier };
+export type SuggestModel = {
+  id: string;
+  label: string;
+  tier: ModelTier;
+  /**
+   * Context window in tokens, from the gateway's catalog for the same model. Only
+   * the gateway can be asked at run time (readGatewayCatalog), so a provider used
+   * directly reads its window here instead of guessing. Null where the gateway
+   * does not carry the model: the run falls back to `BOT_RUN.compactAt`.
+   */
+  context?: number | null;
+};
+
+/**
+ * A model's context window where the app can know it: the gateway's catalog row
+ * for a gateway model, else the window stamped on a directly used provider's
+ * shelf (`SuggestModel.context`). Null when neither says. Pure, so the settings
+ * screen and a run (ai/model compactBudget) read the same number.
+ */
+export function contextWindowOf(
+  provider: TextModelProviderId,
+  model: string,
+  catalog: readonly GatewayModel[] = [],
+): number | null {
+  if (provider === "vercel-ai-gateway") {
+    return catalog.find((row) => row.id === model)?.contextWindow ?? null;
+  }
+  const shelf = TEXT_MODEL_PROVIDERS[provider].suggestModels;
+  return shelf.find((row) => row.id === model)?.context ?? null;
+}
+
+/**
+ * Where a run on a model with this window summarises itself, in tokens:
+ * `BOT_RUN.compactHeadroom` of the window, or `BOT_RUN.compactAt` when the window
+ * is unknown. What a bot's settings fill in, and what a run falls back to.
+ */
+export const compactAtFor = (window: number | null | undefined): number =>
+  window && window > 0
+    ? Math.round(window * BOT_RUN.compactHeadroom)
+    : BOT_RUN.compactAt;
 
 /** Providers that make non-text media. Kept apart from the text providers: the key is the same, but a text provider does not necessarily draw (Anthropic). */
 export const mediaModelProviderSchema = z.enum([
@@ -72,7 +113,7 @@ export type MediaModelRef = z.infer<typeof mediaModelRefSchema>;
  * Suggested media models per provider and kind, cheapest first; the head runs when only the
  * provider is known. Suggestions, not a validation list. Ids follow each sdk's own spelling, so
  * the gateway names the same model differently (`spacexai/…`). An empty list means the provider
- * cannot make that kind (OpenAI has no video model).
+ * cannot make that kind — for a model the sdk cannot build as well as one that does not exist.
  */
 export const MEDIA_MODEL_PROVIDERS: Record<
   MediaModelProviderId,
@@ -88,9 +129,18 @@ export const MEDIA_MODEL_PROVIDERS: Record<
     models: {
       image: [
         { id: "gpt-image-1-mini", label: "GPT Image 1 mini", tier: "small" },
-        { id: "gpt-image-1.5", label: "GPT Image 1.5", tier: "mid" },
-        { id: "gpt-image-2", label: "GPT Image 2", tier: "large" },
+        {
+          id: "gpt-image-2.5-flare",
+          label: "GPT Image 2.5 Flare",
+          tier: "mid",
+        },
+        {
+          id: "gpt-image-2.5-sunburst",
+          label: "GPT Image 2.5 Sunburst",
+          tier: "large",
+        },
       ],
+      // Sora exists, but the sdk's OpenAI provider has no video model (model.ts buildVideoModel)
       video: [],
       speech: [
         { id: "tts-1", label: "TTS 1", tier: "small" },
@@ -98,13 +148,12 @@ export const MEDIA_MODEL_PROVIDERS: Record<
         { id: "tts-1-hd", label: "TTS 1 HD", tier: "large" },
       ],
       transcription: [
-        { id: "whisper-1", label: "Whisper", tier: "small" },
         {
           id: "gpt-4o-mini-transcribe",
           label: "4o mini Transcribe",
           tier: "small",
         },
-        { id: "gpt-4o-transcribe", label: "4o Transcribe", tier: "mid" },
+        { id: "gpt-transcribe", label: "GPT Transcribe", tier: "mid" },
       ],
     },
   },
@@ -114,17 +163,17 @@ export const MEDIA_MODEL_PROVIDERS: Record<
     models: {
       image: [
         {
-          id: "gemini-2.5-flash-image",
-          label: "2.5 Flash Image",
+          id: "gemini-3.1-flash-lite-image",
+          label: "3.1 Flash Lite Image",
           tier: "small",
         },
         {
-          id: "gemini-3.1-flash-image-preview",
+          id: "gemini-3.1-flash-image",
           label: "3.1 Flash Image",
           tier: "mid",
         },
         {
-          id: "gemini-3-pro-image-preview",
+          id: "gemini-3-pro-image",
           label: "3 Pro Image",
           tier: "large",
         },
@@ -140,7 +189,7 @@ export const MEDIA_MODEL_PROVIDERS: Record<
           label: "Veo 3.1 Fast",
           tier: "mid",
         },
-        { id: "veo-3.1-generate", label: "Veo 3.1", tier: "large" },
+        { id: "veo-3.1-generate-preview", label: "Veo 3.1", tier: "large" },
       ],
       speech: [
         {
@@ -175,8 +224,8 @@ export const MEDIA_MODEL_PROVIDERS: Record<
       image: [
         { id: "grok-imagine-image", label: "Grok Imagine", tier: "small" },
         {
-          id: "grok-imagine-image-pro",
-          label: "Grok Imagine Pro",
+          id: "grok-imagine-image-2.0",
+          label: "Grok Imagine 2.0",
           tier: "mid",
         },
       ],
@@ -230,7 +279,11 @@ export const MEDIA_MODEL_PROVIDERS: Record<
           label: "Gemini 3 Pro Image",
           tier: "large",
         },
-        { id: "openai/gpt-image-2", label: "GPT Image 2", tier: "large" },
+        {
+          id: "openai/gpt-image-2.5-sunburst",
+          label: "GPT Image 2.5 Sunburst",
+          tier: "large",
+        },
       ],
       video: [
         {
@@ -344,13 +397,15 @@ export function parseMediaModel(
 
 /**
  * Suggested text models per provider, cheapest first. Suggestions, not a validation list;
- * `defaultModelOf` and `cheapestModelOf` read tiers, not positions.
+ * `defaultModelOf` reads tiers, not positions.
  */
 export const TEXT_MODEL_PROVIDERS: Record<
   TextModelProviderId,
   {
     label: string;
     apiKeyName: string;
+    /** Reached by signing in to an account rather than a pasted key; `apiKeyName` is where the sign-in is kept. */
+    signIn?: true;
     suggestModels: SuggestModel[];
   }
 > = {
@@ -358,19 +413,85 @@ export const TEXT_MODEL_PROVIDERS: Record<
     label: "Open AI",
     apiKeyName: "OPENAI_API_KEY",
     suggestModels: [
-      { id: "gpt-5.6-luna", label: "5.6 Luna", tier: "small" },
-      { id: "gpt-5.6-terra", label: "5.6 Terra", tier: "mid" },
-      { id: "gpt-5.6-sol", label: "5.6 Sol", tier: "large" },
-      { id: "gpt-6-astra", label: "6 Astra", tier: "large" },
+      {
+        id: "gpt-5.6-luna",
+        label: "5.6 Luna",
+        tier: "small",
+        context: 1_050_000,
+      },
+      {
+        id: "gpt-5.6-terra",
+        label: "5.6 Terra",
+        tier: "mid",
+        context: 1_050_000,
+      },
+      {
+        id: "gpt-5.6-sol",
+        label: "5.6 Sol",
+        tier: "large",
+        context: 1_050_000,
+      },
+      {
+        id: "gpt-6-astra",
+        label: "6 Astra",
+        tier: "large",
+        context: 1_050_000,
+      },
+    ],
+  },
+  /**
+   * The Codex models a ChatGPT plan carries, reached by signing in (ai/chatgpt). Which ones a
+   * plan opens differs — a Free plan lists Luna, Terra and 5.5 — and the backend caps every
+   * window at 272k, whatever the same model takes over the API.
+   */
+  chatgpt: {
+    label: "GPT Subscription",
+    apiKeyName: "CHATGPT_SIGN_IN",
+    signIn: true,
+    suggestModels: [
+      {
+        id: "gpt-5.6-luna",
+        label: "5.6 Luna",
+        tier: "small",
+        context: 272_000,
+      },
+      {
+        id: "gpt-5.6-terra",
+        label: "5.6 Terra",
+        tier: "mid",
+        context: 272_000,
+      },
+      {
+        id: "gpt-5.6-sol",
+        label: "5.6 Sol",
+        tier: "large",
+        context: 272_000,
+      },
+      {
+        id: "gpt-6-astra",
+        label: "6 Astra",
+        tier: "large",
+        context: 272_000,
+      },
     ],
   },
   anthropic: {
     label: "Claude",
     apiKeyName: "ANTHROPIC_API_KEY",
     suggestModels: [
-      { id: "claude-haiku-4-5-20251001", label: "Haiku 4.5", tier: "small" },
-      { id: "claude-sonnet-5", label: "Sonnet 5", tier: "mid" },
-      { id: "claude-opus-5", label: "Opus 5", tier: "large" },
+      { id: "claude-haiku-4-5", label: "Haiku 4.5", tier: "small" },
+      {
+        id: "claude-sonnet-5",
+        label: "Sonnet 5",
+        tier: "mid",
+        context: 1_000_000,
+      },
+      {
+        id: "claude-opus-5",
+        label: "Opus 5",
+        tier: "large",
+        context: 1_000_000,
+      },
       { id: "claude-fable-5-1", label: "Fable 5.1", tier: "large" },
     ],
   },
@@ -378,13 +499,26 @@ export const TEXT_MODEL_PROVIDERS: Record<
     label: "Gemini",
     apiKeyName: "GOOGLE_GENERATIVE_AI_API_KEY",
     suggestModels: [
-      { id: "gemini-3.8-flash", label: "3.8 Flash", tier: "mid" },
+      {
+        id: "gemini-3.5-flash-lite",
+        label: "3.5 Flash Lite",
+        tier: "small",
+        context: 1_000_000,
+      },
+      {
+        id: "gemini-3.8-flash",
+        label: "3.8 Flash",
+        tier: "mid",
+        context: 1_000_000,
+      },
     ],
   },
   xai: {
     label: "xAI",
     apiKeyName: "XAI_API_KEY",
-    suggestModels: [{ id: "grok-4.6", label: "Grok 4.6", tier: "large" }],
+    suggestModels: [
+      { id: "grok-4.6", label: "Grok 4.6", tier: "large", context: 500_000 },
+    ],
   },
   /**
    * Cross-vendor shelf, cheapest first, shown until the live catalog arrives. Vendor names are
@@ -400,24 +534,55 @@ export const TEXT_MODEL_PROVIDERS: Record<
         label: "Gemini 3.8 Flash",
         tier: "small",
       },
-      { id: "openai/gpt-5.6-luna", label: "GPT 5.6 Luna", tier: "small" },
+      {
+        id: "openai/gpt-5.6-luna",
+        label: "GPT 5.6 Luna",
+        tier: "small",
+        context: 1_050_000,
+      },
       { id: "zai/glm-5.3", label: "GLM 5.3", tier: "mid" },
-      { id: "spacexai/grok-4.6", label: "Grok 4.6", tier: "mid" },
+      {
+        id: "spacexai/grok-4.6",
+        label: "Grok 4.6",
+        tier: "mid",
+        context: 500_000,
+      },
       {
         id: "anthropic/claude-sonnet-5",
         label: "Claude Sonnet 5",
         tier: "mid",
       },
-      { id: "openai/gpt-5.6-terra", label: "GPT 5.6 Terra", tier: "mid" },
+      {
+        id: "openai/gpt-5.6-terra",
+        label: "GPT 5.6 Terra",
+        tier: "mid",
+        context: 1_050_000,
+      },
       { id: "moonshotai/kimi-k3", label: "Kimi K3", tier: "mid" },
-      { id: "openai/gpt-5.6-sol", label: "GPT 5.6 Sol", tier: "large" },
-      { id: "anthropic/claude-opus-5", label: "Claude Opus 5", tier: "large" },
+      {
+        id: "openai/gpt-5.6-sol",
+        label: "GPT 5.6 Sol",
+        tier: "large",
+        context: 1_050_000,
+      },
+      {
+        id: "anthropic/claude-opus-5",
+        label: "Claude Opus 5",
+        tier: "large",
+        context: 1_000_000,
+      },
+      {
+        id: "openai/gpt-6-astra",
+        label: "GPT 6 Astra",
+        tier: "large",
+        context: 1_050_000,
+      },
     ],
   },
 };
 
 /** A provider's model at one tier, or its first one when it has none at that size. */
-export const modelOfTier = (
+const modelOfTier = (
   provider: { suggestModels: SuggestModel[] },
   tier: ModelTier,
 ): string | undefined =>
@@ -430,10 +595,6 @@ export const modelOfTier = (
 export const defaultModelOf = (provider: { suggestModels: SuggestModel[] }) =>
   modelOfTier(provider, "mid");
 
-/** The cheapest model of a provider; what `web_search` runs on (model.ts resolveSearchModel). */
-export const cheapestModelOf = (provider: { suggestModels: SuggestModel[] }) =>
-  modelOfTier(provider, "small");
-
 /**
  * The speech-to-speech backends. `models` and `voices` are suggestion lists, not validation
  * lists (the api takes more, e.g. a cloned voice id); the first model is the default. Written
@@ -536,8 +697,30 @@ export type AiProvider = {
   apiKeyName: string;
   hasKey: boolean;
   suggestModels: SuggestModel[];
+  /** Signs in to an account instead of taking a key (TEXT_MODEL_PROVIDERS `signIn`). */
+  signIn?: true;
+  /** The plan the signed-in account is on, as the provider names it; sign-in providers only. */
+  plan?: string | null;
 };
 
+/**
+ * What a GPT Subscription sign-in has used (ai/chatgpt readChatGptUsage): the tightest window of
+ * its plan and when it frees up, with `plan` as the backend names it now. `refused` is a sign-in
+ * the plan turned away, in its own words.
+ */
+export type SubscriptionUsage =
+  | {
+      plan: string | null;
+      usedPercent: number;
+      /** ISO; null when the backend did not say. */
+      resetsAt: string | null;
+      /** The window is used up: jobs on it stop until it resets. */
+      spent: boolean;
+      /** Spent, or past `CHATGPT_USAGE_HIGH`: amber on its row. */
+      high: boolean;
+    }
+  | { refused: string };
+
 /** What a gateway row can be picked as: text, one of the studio kinds, or a word this app offers nowhere (embedding, reranking, realtime). */
 export const GATEWAY_TEXT = "language";
 
@@ -571,4 +754,19 @@ export type GatewayModel = {
   price: GatewayPrice;
   /** The gateway has dated its retirement. */
   retiring: boolean;
+  /** Context window in tokens; null when the gateway did not say. What a run compacts against (bot.run). */
+  contextWindow: number | null;
 };
+
+/**
+ * What is left on the gateway key (ai/model `readGatewayCredits`). A key the gateway turns
+ * away is not a failed read: `refused` carries its words, drawn on the key's row and dialog.
+ */
+export type GatewayCredits =
+  | {
+      /** USD. */
+      balance: number;
+      /** At or under `GATEWAY_LOW_CREDIT` (config): amber, it waits on a top-up. */
+      low: boolean;
+    }
+  | { refused: string };

package/features/ai/model.ts

@@ -8,10 +8,13 @@ import {
   type experimental_generateVideo,
   type ImageModel,
   type LanguageModel,
+  NoOutputGeneratedError,
+  RetryError,
   type SpeechModel,
   type ToolSet,
   type TranscriptionModel,
 } from "ai";
+import { GATEWAY_LOW_CREDIT } from "@/config";
 import {
   DEFAULT_MODEL_KEY,
   MEDIA_MODEL_KEYS,
@@ -20,11 +23,14 @@ import { readConfig } from "@/features/config/config.query";
 import { logger } from "@/lib/logger";
 import { publicError } from "@/lib/public-error";
 import { clip, errorToString } from "@/lib/utils";
+import { chatGptModel } from "./chatgpt";
 import {
   canMakeKind,
-  cheapestModelOf,
+  compactAtFor,
+  contextWindowOf,
   defaultModelOf,
   GATEWAY_TEXT,
+  type GatewayCredits,
   type GatewayModel,
   type GatewayPrice,
   MEDIA_MODEL_PROVIDERS,
@@ -67,6 +73,92 @@ export function modelErrorToString(cause: unknown): string {
   return `${cause.message}${status}${said}`;
 }
 
+/**
+ * What a failed model call means for the run that made it (bot.runner): `transient`
+ * is trouble a moment fixes — a rate limit, an overload, a dropped connection, a call
+ * that went quiet; `overflow` is a context the model refused as too long, which a
+ * compaction fixes; anything else is `fatal`, a person's to act on (a key, the credit,
+ * a model id). The sdk wraps retries and causes, so the whole chain is read.
+ */
+export type ModelFailure = "transient" | "overflow" | "fatal";
+
+/** Connection failures by the code Node or undici gives them. */
+const NETWORK_CODES = new Set([
+  "ECONNRESET",
+  "ECONNREFUSED",
+  "ETIMEDOUT",
+  "EPIPE",
+  "ENOTFOUND",
+  "EAI_AGAIN",
+  "ENETUNREACH",
+  "EHOSTUNREACH",
+  "UND_ERR_SOCKET",
+  "UND_ERR_CLOSED",
+  "UND_ERR_CONNECT_TIMEOUT",
+  "UND_ERR_HEADERS_TIMEOUT",
+  "UND_ERR_BODY_TIMEOUT",
+]);
+
+/**
+ * How providers word a context too long for the model; the sdk has no error class for
+ * it. Read off a 400 or a 413, or off an error with no status of its own — a retry
+ * wrapper repeats the words of the one it wraps.
+ */
+const TOO_LONG =
+  /context[ _-]?(length|window)|maximum context|prompt is too long|too many (input )?tokens|input token count|exceeds? the (maximum|model'?s?) (context|input|prompt|number of tokens)|maximum prompt length|request (entity )?too large/i;
+
+export function modelFailureOf(cause: unknown): ModelFailure {
+  let transient = false;
+  for (const error of causeChain(cause)) {
+    const { statusCode, isRetryable, code, name, message } = error as {
+      statusCode?: unknown;
+      isRetryable?: unknown;
+      code?: unknown;
+      name?: unknown;
+      message?: unknown;
+    };
+    const body = APICallError.isInstance(error)
+      ? (error.responseBody ?? "")
+      : "";
+    if (
+      (statusCode === undefined || statusCode === 400 || statusCode === 413) &&
+      TOO_LONG.test(`${String(message ?? "")} ${body}`)
+    ) {
+      return "overflow";
+    }
+    if (
+      isRetryable === true ||
+      name === "TimeoutError" ||
+      NoOutputGeneratedError.isInstance(error) ||
+      (typeof statusCode === "number" &&
+        (statusCode === 408 ||
+          statusCode === 409 ||
+          statusCode === 429 ||
+          statusCode >= 500)) ||
+      (typeof code === "string" && NETWORK_CODES.has(code))
+    ) {
+      transient = true;
+    }
+  }
+  return transient ? "transient" : "fatal";
+}
+
+/** An error and everything it wraps — its `cause`, each attempt of a retry. */
+function causeChain(cause: unknown): object[] {
+  const chain: object[] = [];
+  const queue: unknown[] = [cause];
+  while (queue.length && chain.length < 12) {
+    const next = queue.shift();
+    if (typeof next !== "object" || next === null || chain.includes(next)) {
+      continue;
+    }
+    chain.push(next);
+    if (RetryError.isInstance(next)) queue.push(...next.errors);
+    if ("cause" in next) queue.push(next.cause);
+  }
+  return chain;
+}
+
 /** A resolved model. Provider-native tools (web search) are built off the provider instance, not the model, so both come out of one switch. */
 export type TextModel = {
   ref: TextModelRef;
@@ -86,6 +178,9 @@ export function buildTextModel(ref: TextModelRef, apiKey: string): TextModel {
         searchTools: { search: openai.tools.webSearch() },
       };
     }
+    case "chatgpt":
+      // Signs every request with the stored sign-in itself (ai/chatgpt), so no key is passed
+      return { ref, model: chatGptModel(ref.model), searchTools: null };
     case "anthropic": {
       const anthropic = createAnthropic({ apiKey });
       return {
@@ -143,6 +238,7 @@ type CatalogRow = {
   deprecated_at?: number | null;
   modalities?: { output?: string[] | null } | null;
   pricing?: Record<string, unknown> | null;
+  context_window?: number | null;
 };
 
 /** USD per 1M tokens from the gateway's per-token string. Null is "it did not say", never zero. */
@@ -244,6 +340,10 @@ export async function readGatewayCatalog(): Promise<GatewayModel[]> {
         tags: row.tags ?? [],
         price: priceOfGatewayModel(row),
         retiring: Boolean(row.deprecated_at),
+        contextWindow:
+          typeof row.context_window === "number" && row.context_window > 0
+            ? row.context_window
+            : null,
       }),
     )
     .sort((a, b) => a.id.localeCompare(b.id));
@@ -252,6 +352,66 @@ export async function readGatewayCatalog(): Promise<GatewayModel[]> {
   return list;
 }
 
+/** Unlike the listing, this answers only to a key. */
+const GATEWAY_CREDITS_URL = "https://ai-gateway.vercel.sh/v1/credits";
+
+/**
+ * What is left on the gateway key. The gateway is the one provider that tells this to the key
+ * a user typed; the others keep a balance behind an admin key. A key it turns away is a state
+ * of that key, not a failed read, so it comes back as `refused` in the gateway's own words.
+ */
+export async function readGatewayCredits(): Promise<GatewayCredits | null> {
+  const apiKey = await readConfig(
+    TEXT_MODEL_PROVIDERS["vercel-ai-gateway"].apiKeyName,
+  );
+  if (!apiKey) return null;
+
+  const response = await fetch(GATEWAY_CREDITS_URL, {
+    headers: { accept: "application/json", authorization: `Bearer ${apiKey}` },
+  }).catch((cause: unknown) => {
+    logger.warn({ cause }, "gateway credits unreachable");
+    publicError("Could not reach the gateway");
+  });
+
+  const body = (await response.json().catch(() => null)) as {
+    balance?: string;
+    error?: { message?: string };
+  } | null;
+  const said = clip(body?.error?.message ?? "", PROVIDER_BODY_MAX);
+
+  if (response.status === 401 || response.status === 403)
+    return { refused: said || `The gateway answered ${response.status}` };
+  if (!response.ok)
+    publicError(
+      `The gateway answered ${response.status}${said ? `: ${said}` : ""}`,
+    );
+
+  const balance = Number(body?.balance);
+  if (!body?.balance || !Number.isFinite(balance))
+    publicError("The gateway did not say what is left");
+  return { balance, low: balance <= GATEWAY_LOW_CREDIT };
+}
+
+/**
+ * Where a run summarises itself, in tokens (bot.run): what the bot's owner set,
+ * else the model's own window worked out the same way the settings screen does
+ * (model.schema compactAtFor). An unknown window falls back to `BOT_RUN.compactAt`,
+ * which assumes the window a current model carries.
+ */
+export async function compactBudget(
+  ref: TextModelRef,
+  /** What the bot's owner set, if anything (bot.compactAt); it wins. */
+  chosen?: number | null,
+): Promise<number> {
+  if (chosen && chosen > 0) return chosen;
+  // Never worth failing a run over: an unreachable catalog is a fallback, not an error
+  const catalog =
+    ref.provider === "vercel-ai-gateway"
+      ? await readGatewayCatalog().catch(() => [])
+      : [];
+  return compactAtFor(contextWindowOf(ref.provider, ref.model, catalog));
+}
+
 /**
  * The gateway shelf as a set of ids. Null means it could not be asked, which is not "not there":
  * callers keep the written rows. Only existence is checked; the gateway prices image and video
@@ -280,9 +440,15 @@ async function callableRows<T extends { id: string }>(
 }
 
 export async function getTextModel(ref: TextModelRef): Promise<TextModel> {
-  const { label, apiKeyName } = TEXT_MODEL_PROVIDERS[ref.provider];
+  const { label, apiKeyName, signIn } = TEXT_MODEL_PROVIDERS[ref.provider];
   const apiKey = await readConfig(apiKeyName);
-  if (!apiKey) publicError(`No ${label} key — add one in Config.`);
+  if (!apiKey) {
+    publicError(
+      signIn
+        ? `${label} is not signed in — sign in from Config.`
+        : `No ${label} key — add one in Config.`,
+    );
+  }
   return buildTextModel(ref, apiKey);
 }
 
@@ -340,27 +506,6 @@ export async function resolveDefaultModel(
   publicError("No model key is set — add one in Config");
 }
 
-/**
- * The model a search runs on: this run's own when it has a native search, else the first
- * search-capable provider with a key, else null (the tool drops off the list).
- */
-export async function resolveSearchModel(
-  run?: TextModel | null,
-): Promise<TextModel | null> {
-  if (run?.searchTools) return run;
-
-  for (const provider of TEXT_MODEL_PROVIDER_LIST) {
-    const apiKey = await readConfig(provider.apiKeyName);
-    // The smallest one it has: this call reads pages and hands back a
-    // paragraph, and nobody upstream can tell what it ran on
-    const model = cheapestModelOf(provider);
-    if (!apiKey || !model) continue;
-    const candidate = buildTextModel({ provider: provider.id, model }, apiKey);
-    if (candidate.searchTools) return candidate;
-  }
-  return null;
-}
-
 /** The sdk declares this one but does not export it — read off the function that takes it. */
 type VideoModel = Parameters<typeof experimental_generateVideo>[0]["model"];