thursday-agent 0.4.0 → 0.6.0

package/features/ai/model.ts

@@ -8,7 +8,6 @@ import {
   type experimental_generateVideo,
   type ImageModel,
   type LanguageModel,
-  NoOutputGeneratedError,
   RetryError,
   type SpeechModel,
   type ToolSet,
@@ -29,7 +28,6 @@ import {
   compactAtFor,
   contextWindowOf,
   defaultModelOf,
-  GATEWAY_TEXT,
   type GatewayCredits,
   type GatewayModel,
   type GatewayPrice,
@@ -39,7 +37,6 @@ import {
   type MediaModelRef,
   parseMediaModel,
   parseTextModel,
-  SPEACH_MODEL_PROVIDER_LIST,
   type SuggestModel,
   TEXT_MODEL_PROVIDER_LIST,
   TEXT_MODEL_PROVIDERS,
@@ -73,74 +70,53 @@ export function modelErrorToString(cause: unknown): string {
   return `${cause.message}${status}${said}`;
 }
 
+/** How providers word a context too long for the model; the sdk has no error class for it. */
+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;
+
 /**
- * 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.
+ * Whether the model refused the context as too long, which a compaction fixes (whether
+ * anything else may pass on another try is `isProviderRefusal`). Read off the words of
+ * a 400, a 413, or an error with no status of its own, anywhere in the chain: a retry
+ * wrapper repeats the words of the one it wraps, and a streamed body the sdk could not
+ * read as an error keeps them in a bare `error` string.
  */
-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",
-]);
+export function isContextOverflow(cause: unknown): boolean {
+  return causeChain(cause).some((error) => {
+    const {
+      statusCode,
+      message,
+      error: inner,
+    } = error as { statusCode?: unknown; message?: unknown; error?: unknown };
+    if (statusCode !== undefined && statusCode !== 400 && statusCode !== 413)
+      return false;
+    const body = APICallError.isInstance(error) ? error.responseBody : null;
+    return [message, inner, body].some(
+      (words) => typeof words === "string" && TOO_LONG.test(words),
+    );
+  });
+}
+
+/** 4xx statuses that are a moment's trouble, not a refusal: a timeout, a conflict, a rate limit. */
+const PASSING_4XX = new Set([408, 409, 429]);
 
 /**
- * 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.
+ * Whether the provider refused the call — the key, the credit, the model id, a request
+ * it will never take — which only a person can change. Anything else broke on the way
+ * (a dropped or garbled stream, an overload, a model gone quiet) and may pass on another
+ * try, and so may a context refused as too long, once it is compacted.
  */
-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";
+export function isProviderRefusal(cause: unknown): boolean {
+  if (isContextOverflow(cause)) return false;
+  return causeChain(cause).some((error) => {
+    const { statusCode } = error as { statusCode?: unknown };
+    return (
+      typeof statusCode === "number" &&
+      statusCode >= 400 &&
+      statusCode < 500 &&
+      !PASSING_4XX.has(statusCode)
+    );
+  });
 }
 
 /** An error and everything it wraps — its `cause`, each attempt of a retry. */
@@ -236,7 +212,6 @@ type CatalogRow = {
   type?: string | null;
   tags?: string[] | null;
   deprecated_at?: number | null;
-  modalities?: { output?: string[] | null } | null;
   pricing?: Record<string, unknown> | null;
   context_window?: number | null;
 };
@@ -248,18 +223,6 @@ function per1M(value: unknown): number | null {
   return Number.isFinite(amount) ? Math.round(amount * 1e6 * 1e4) / 1e4 : null;
 }
 
-/**
- * What a row may be offered as. The gateway's own word is right except for the models that
- * both talk and draw — Gemini's `*-image` family answers `language`, and a picker that
- * believed it would leave them out of the image list. Their output modality gives them away.
- */
-function kindOfGatewayModel(row: CatalogRow) {
-  const type = row.type ?? null;
-  if (type === GATEWAY_TEXT && (row.modalities?.output ?? []).includes("image"))
-    return "image";
-  return type;
-}
-
 /**
  * The gateway bills ten different ways; this is the one place that knows it. Billed by the
  * second or the character, a token price would misstate the model, so the unit replaces it;
@@ -313,7 +276,7 @@ function priceOfGatewayModel(row: CatalogRow): GatewayPrice {
 
 /**
  * What the gateway carries now, every modality in one listing. Rows come back with their kind
- * settled (`kindOfGatewayModel`) and their price flattened (`priceOfGatewayModel`), so a screen
+ * as the gateway states it and their price flattened (`priceOfGatewayModel`), so a screen
  * filters and sorts on plain fields. The gateway is the only provider that can be asked.
  */
 export async function readGatewayCatalog(): Promise<GatewayModel[]> {
@@ -336,7 +299,10 @@ export async function readGatewayCatalog(): Promise<GatewayModel[]> {
         id: row.id,
         label: row.name || row.id,
         owner: row.owned_by || row.id.slice(0, row.id.indexOf("/")),
-        type: kindOfGatewayModel(row),
+        // The gateway's own word. A Gemini `*-image` row says `language` and lists
+        // `text,image` output, but the gateway's image endpoint takes `type: "image"`
+        // only, so offering one as an image model only makes a choice that always fails.
+        type: row.type ?? null,
         tags: row.tags ?? [],
         price: priceOfGatewayModel(row),
         retiring: Boolean(row.deprecated_at),
@@ -455,8 +421,7 @@ export async function getTextModel(ref: TextModelRef): Promise<TextModel> {
 /**
  * The model when nobody picked one; a bot with no model asks here on every run (bot.run
  * resolveModel). Order: the named provider, then the default set in Settings > Models, then
- * the provider the calls run on (a voice key is also a text key, model.schema
- * SPEACH_MODEL_PROVIDERS), then any provider with a key.
+ * OpenAI and xAI, then any provider with a key.
  */
 export async function resolveDefaultModel(
   providerId?: string,
@@ -472,10 +437,9 @@ export async function resolveDefaultModel(
   const named = providerId
     ? TEXT_MODEL_PROVIDER_LIST.find((v) => v.id === providerId)
     : undefined;
-  // The provider the calls run on; REALTIME_PROVIDERS is a subset of the text list, so ids carry over (realtime.schema)
-  const voiced = SPEACH_MODEL_PROVIDER_LIST.map((provider) => ({
-    ...TEXT_MODEL_PROVIDERS[provider.id],
-    id: provider.id,
+  const preferred = (["openai", "xai"] as const).map((id) => ({
+    ...TEXT_MODEL_PROVIDERS[id],
+    id,
   }));
 
   if (named) {
@@ -493,7 +457,7 @@ export async function resolveDefaultModel(
   )
     return chosen;
 
-  for (const provider of [...voiced, ...TEXT_MODEL_PROVIDER_LIST]) {
+  for (const provider of [...preferred, ...TEXT_MODEL_PROVIDER_LIST]) {
     if (!(await readConfig(provider.apiKeyName))) continue;
     const model = await workhorse(provider);
     if (model) {