@providerkit/core 0.8.0 → 0.9.0

package/src/fallback.ts

@@ -1,6 +1,8 @@
 import { classify, parseRetryAfterMs, ProviderError, type ErrorKind } from "./errors.ts";
 import type { RateLimitWindow } from "./rate-limit.ts";
+import type { ProviderPresetId } from "./presets.ts";
 import type { Provider, ProviderChunk } from "./types.ts";
+import { createPresetProvider, type PresetProviderConfig } from "./providers/factory.ts";
 
 /** Retry intervals when the server supplies no deadline, not estimates of when
  *  it will recover. A server deadline always wins over these defaults. */
@@ -30,10 +32,31 @@ export interface FallbackOptions<T> {
   now?: () => number;
 }
 
+/**
+ * A fallback named by PRESET id rather than hand-built: the endpoint, auth
+ * style, effort dialect and every measured quirk come from the presets table,
+ * so a consumer writes neither a baseUrl nor a model-spelling rule. A
+ * hand-built `Provider` is still accepted wherever a spec is.
+ */
+export interface FallbackSpec extends PresetProviderConfig {
+  preset: ProviderPresetId;
+}
+
+export type FallbackCandidate = Provider | FallbackSpec;
+
+function resolveFallback(candidate: FallbackCandidate): Provider {
+  if (typeof candidate === "object" && candidate !== null && "createStream" in candidate) {
+    return candidate;
+  }
+  const { preset, ...config } = candidate;
+  return createPresetProvider(preset, config);
+}
+
 /** Standard options that any provider config can accept to configure fallbacks inline. */
 export interface ProviderFallbackConfig {
-  /** Secondary and tertiary providers to try if this provider fails/exhausts. */
-  fallbacks?: readonly Provider[];
+  /** Secondary and tertiary providers to try if this provider fails/exhausts:
+   *  hand-built `Provider`s, preset-id `FallbackSpec`s, or both. */
+  fallbacks?: readonly FallbackCandidate[];
   /** Custom cooldown or telemetry options for the fallback pool. */
   fallbackOptions?: FallbackOptions<Provider>;
 }
@@ -47,7 +70,10 @@ export function withConfiguredFallbacks(
   config?: ProviderFallbackConfig,
 ): Provider {
   if (!config?.fallbacks || config.fallbacks.length === 0) return primary;
-  return withFallbackProviders([primary, ...config.fallbacks], config.fallbackOptions);
+  return withFallbackProviders(
+    [primary, ...config.fallbacks.map(resolveFallback)],
+    config.fallbackOptions,
+  );
 }
 
 interface Cooldown {

package/src/presets.ts

@@ -37,9 +37,11 @@ export type PresetAuth = "key" | "bearer" | "oauth";
 export interface ProviderPreset {
   /** Adapter wire format. */
   shape: PresetShape;
-  /** API root — the adapter appends its own path (/v1/chat/completions,
-   *  /v1/messages, …), so this must NOT carry the version. */
+  /** API root — if it ends with /v1, /v4 or another version segment, the
+   *  adapter automatically routes to /chat/completions instead of /v1/chat/completions. */
   baseUrl: string;
+  /** Explicit path override when a gateway has custom routing. */
+  path?: string;
   auth: PresetAuth;
   /** Fallback when the caller passes no model. */
   defaultModel?: string;
@@ -65,15 +67,15 @@ export const PROVIDER_PRESETS = {
     shape: "anthropic",
     baseUrl: "https://api.anthropic.com",
     auth: "key",
-    defaultModel: "claude-sonnet-4-6",
-    models: ["claude-sonnet-4-6", "claude-opus-5", "claude-fable-5-1"],
+    defaultModel: "claude-sonnet-5",
+    models: ["claude-sonnet-5", "claude-opus-5", "claude-fable-5-1", "claude-haiku-4-5"],
   },
   openai: {
     shape: "openai",
     baseUrl: "https://api.openai.com",
     auth: "key",
     defaultModel: "gpt-5.6-sol",
-    models: ["gpt-5.6-sol", "gpt-5-pro", "gpt-5-nano"],
+    models: ["gpt-5.6-sol", "gpt-5.5-pro", "gpt-5.4-mini"],
   },
   google: {
     // Native Generative Language REST — the adapter authenticates with
@@ -81,8 +83,8 @@ export const PROVIDER_PRESETS = {
     shape: "gemini",
     baseUrl: "https://generativelanguage.googleapis.com",
     auth: "key",
-    defaultModel: "gemini-3.1-pro",
-    models: ["gemini-3.1-pro", "gemini-2.5-flash"],
+    defaultModel: "gemini-3.1-pro-preview",
+    models: ["gemini-3.1-pro-preview", "gemini-3.8-flash", "gemini-flash-latest"],
   },
   /** Gemini behind its OpenAI-compatible route — the drop-in for codebases
    *  that only speak the openai shape. */
@@ -90,15 +92,17 @@ export const PROVIDER_PRESETS = {
     shape: "openai",
     baseUrl: "https://generativelanguage.googleapis.com/v1beta/openai",
     auth: "key",
-    defaultModel: "gemini-3.1-pro",
-    models: ["gemini-3.1-pro", "gemini-2.5-flash"],
+    defaultModel: "gemini-3.1-pro-preview",
+    models: ["gemini-3.1-pro-preview", "gemini-3.8-flash", "gemini-flash-latest"],
   },
   deepseek: {
     shape: "openai",
     baseUrl: "https://api.deepseek.com",
     auth: "key",
-    defaultModel: "deepseek-v4-flash",
-    models: ["deepseek-v4-flash", "deepseek-v4-pro", "deepseek-flash"],
+    // Measured 2026-09-13 via /models on live key: first-party serves only
+    // deepseek-flash and deepseek-v4-pro. v4.1-flash is OpenRouter/Baseten.
+    defaultModel: "deepseek-flash",
+    models: ["deepseek-flash", "deepseek-v4-pro"],
     // Text-only API — an image part in the body is a hard 400 (measured).
     textOnly: true,
   },
@@ -107,7 +111,7 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://openrouter.ai/api",
     auth: "key",
     defaultModel: "deepseek/deepseek-v4.1-flash",
-    models: ["deepseek/deepseek-v4.1-flash", "z-ai/glm-5.3-flash", "qwen/qwen3.7-max"],
+    models: ["deepseek/deepseek-v4.1-flash", "z-ai/glm-5.3-flash", "x-ai/grok-4.6", "openai/gpt-5.6-luna-pro"],
   },
   groq: {
     shape: "openai",
@@ -121,25 +125,28 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://api.mistral.ai/v1",
     auth: "key",
     defaultModel: "devstral-2512",
-    models: ["devstral-2512", "magistral-small", "mistral-small-2506"],
+    models: ["devstral-2512", "mistral-large-latest", "magistral-small"],
   },
   xai: {
     shape: "openai",
     baseUrl: "https://api.x.ai/v1",
     auth: "key",
-    defaultModel: "grok-4.5",
-    models: ["grok-4.5", "grok-4.3", "grok-4.20-0309-reasoning"],
+    defaultModel: "grok-4.6",
+    models: ["grok-4.6", "grok-4.5", "grok-4.3"],
   },
   together: {
     shape: "openai",
     baseUrl: "https://api.together.xyz/v1",
     auth: "key",
-    defaultModel: "deepseek-ai/DeepSeek-V3",
+    defaultModel: "MiniMaxAI/MiniMax-M3",
+    models: ["MiniMaxAI/MiniMax-M3", "Qwen/Qwen3-Coder-Next-FP8"],
   },
   fireworks: {
     shape: "openai",
     baseUrl: "https://api.fireworks.ai/inference/v1",
     auth: "key",
+    defaultModel: "accounts/fireworks/models/deepseek-v4p1-flash",
+    models: ["accounts/fireworks/models/deepseek-v4p1-flash", "accounts/fireworks/models/glm-5p3-flash"],
   },
   cerebras: {
     shape: "openai",
@@ -158,16 +165,112 @@ export const PROVIDER_PRESETS = {
   zhipu: {
     shape: "openai",
     baseUrl: "https://api.z.ai/api/paas/v4",
+    path: "/chat/completions",
     auth: "key",
     defaultModel: "glm-5.3-flash",
-    models: ["glm-5.3-flash", "glm-5.2", "glm-4.7"],
+    models: ["glm-5.3-flash", "glm-5.3", "glm-5.2"],
   },
   minimax: {
     shape: "anthropic",
     baseUrl: "https://api.minimax.io/anthropic",
     auth: "key",
-    defaultModel: "MiniMax-M2.5",
-    models: ["MiniMax-M3", "MiniMax-M2.5", "MiniMax-M2.1"],
+    defaultModel: "MiniMax-M3",
+    models: ["MiniMax-M3", "MiniMax-M2.7", "MiniMax-M2.5"],
+  },
+
+  // ── Popular gateways & cloud providers ──────────────────────────────────
+  nvidia: {
+    shape: "openai",
+    baseUrl: "https://integrate.api.nvidia.com/v1",
+    auth: "key",
+    defaultModel: "deepseek-ai/deepseek-v4-flash",
+    models: ["deepseek-ai/deepseek-v4-flash", "deepseek-ai/deepseek-v4-pro"],
+  },
+  perplexity: {
+    shape: "openai",
+    baseUrl: "https://api.perplexity.ai",
+    auth: "key",
+    defaultModel: "sonar",
+    models: ["sonar", "sonar-pro", "sonar-reasoning-pro", "sonar-deep-research"],
+  },
+  deepinfra: {
+    shape: "openai",
+    baseUrl: "https://api.deepinfra.com/v1/openai",
+    path: "/chat/completions",
+    auth: "key",
+    defaultModel: "MiniMaxAI/MiniMax-M3",
+    models: ["MiniMaxAI/MiniMax-M3", "ByteDance/Seed-2.0-pro"],
+  },
+  nebius: {
+    shape: "openai",
+    baseUrl: "https://api.tokenfactory.nebius.com/v1",
+    auth: "key",
+    defaultModel: "Qwen/Qwen3.5-397B-A17B",
+    models: ["Qwen/Qwen3.5-397B-A17B", "deepseek-ai/DeepSeek-V4-Flash-0731"],
+  },
+  novita: {
+    shape: "openai",
+    baseUrl: "https://api.novita.ai/openai",
+    path: "/v1/chat/completions",
+    auth: "key",
+    defaultModel: "baichuan/baichuan-m2-32b",
+    models: ["baichuan/baichuan-m2-32b", "baidu/ernie-4.5-300b-a47b-paddle"],
+  },
+  chutes: {
+    shape: "openai",
+    baseUrl: "https://llm.chutes.ai/v1",
+    auth: "key",
+    defaultModel: "Qwen/Qwen3.8-27B-TEE",
+    models: ["Qwen/Qwen3.8-27B-TEE", "Qwen/Qwen3.5-397B-A17B-TEE"],
+  },
+  siliconflow: {
+    shape: "openai",
+    baseUrl: "https://api.siliconflow.com/v1",
+    auth: "key",
+    defaultModel: "Qwen/Qwen3-235B-A22B-Thinking-2507",
+    models: ["Qwen/Qwen3-235B-A22B-Thinking-2507", "MiniMaxAI/MiniMax-M2.5"],
+  },
+  volcengine: {
+    shape: "openai",
+    baseUrl: "https://ark.cn-beijing.volces.com/api/v3",
+    auth: "key",
+    defaultModel: "doubao-seed-1-8-251228",
+    models: ["doubao-seed-1-8-251228", "deepseek-v4-flash-ga-260731"],
+  },
+  stepfun: {
+    shape: "openai",
+    baseUrl: "https://api.stepfun.com/v1",
+    auth: "key",
+    defaultModel: "step-3.7-flash",
+    models: ["step-3.7-flash", "step-3.5-flash"],
+  },
+  digitalocean: {
+    shape: "openai",
+    baseUrl: "https://inference.do-ai.run/v1",
+    auth: "key",
+    defaultModel: "anthropic-claude-3.7-sonnet",
+    models: ["anthropic-claude-3.7-sonnet", "alibaba-qwen3-32b"],
+  },
+  crusoe: {
+    shape: "openai",
+    baseUrl: "https://api.inference.crusoecloud.com/v1",
+    auth: "key",
+    defaultModel: "Qwen/Qwen3-235B-A22B-Instruct-2507",
+    models: ["Qwen/Qwen3-235B-A22B-Instruct-2507", "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B"],
+  },
+  baseten: {
+    shape: "openai",
+    baseUrl: "https://inference.baseten.co/v1",
+    auth: "key",
+    defaultModel: "deepseek-ai/DeepSeek-V4.1-Flash",
+    models: ["deepseek-ai/DeepSeek-V4.1-Flash", "MiniMaxAI/MiniMax-M2.5"],
+  },
+  huggingface: {
+    shape: "openai",
+    baseUrl: "https://router.huggingface.co/v1",
+    auth: "key",
+    defaultModel: "MiniMaxAI/MiniMax-M3",
+    models: ["MiniMaxAI/MiniMax-M3", "Qwen/Qwen2.5-Coder-32B-Instruct"],
   },
 
   // ── Coding plans — Anthropic wire + Bearer (measured family trait) ─────
@@ -176,7 +279,7 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://api.z.ai/api/anthropic",
     auth: "bearer",
     defaultModel: "glm-5.3-flash",
-    models: ["glm-5.3-flash", "glm-5.3-highspeed", "glm-5.2", "glm-5-turbo"],
+    models: ["glm-5.3-flash", "glm-5.3-highspeed", "glm-5.3", "glm-5.2"],
     // Model ids are BARE here — the gateway-prefixed spelling
     // (`z-ai/glm-5.3-flash`) answers 400 [1211] Unknown Model.
     explicitNone: true,
@@ -186,15 +289,15 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://api.kimi.ai/coding",
     auth: "bearer",
     defaultModel: "kimi-for-coding",
-    models: ["kimi-for-coding", "kimi-for-coding-highspeed", "k3"],
+    models: ["kimi-for-coding", "kimi-for-coding-highspeed", "k3", "k3-256k"],
   },
   /** Alibaba's dashscope coding plan — OpenAI-compatible per its registry. */
   "alibaba-coding-plan": {
     shape: "openai",
     baseUrl: "https://coding-intl.dashscope.aliyuncs.com/v1",
     auth: "bearer",
-    defaultModel: "qwen3.7-max",
-    models: ["qwen3.7-max", "qwen3-coder-next", "glm-4.7"],
+    defaultModel: "qwen3.5-plus",
+    models: ["qwen3.5-plus", "qwen3-coder-next", "glm-5"],
   },
   /** QwenCloud's token plan — Anthropic wire (measured by tabrunner). */
   qwen: {
@@ -202,7 +305,7 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic",
     auth: "bearer",
     defaultModel: "qwen3.8-max",
-    models: ["qwen3.8-max", "qwen3.6-flash"],
+    models: ["qwen3.8-max", "qwen3.8-flash", "qwen3.6-flash"],
   },
 
   // ── Local runtimes ──────────────────────────────────────────────────────
@@ -227,8 +330,8 @@ export const PROVIDER_PRESETS = {
     shape: "anthropic",
     baseUrl: "https://api.anthropic.com",
     auth: "oauth",
-    defaultModel: "claude-sonnet-4-6",
-    models: ["claude-sonnet-4-6", "claude-opus-5"],
+    defaultModel: "claude-sonnet-5",
+    models: ["claude-sonnet-5", "claude-opus-5", "claude-fable-5"],
     // The beta flag is what switches the API into OAuth-token mode (measured
     // by tabrunner's Claude plan sign-in).
     headers: { "anthropic-beta": "claude-code-20250219,oauth-2025-04-20" },
@@ -240,7 +343,7 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://chatgpt.com/backend-api/codex",
     auth: "oauth",
     defaultModel: "gpt-5.3-codex",
-    models: ["gpt-5.3-codex", "gpt-5.5", "gpt-5.4-mini"],
+    models: ["gpt-5.3-codex", "gpt-5.3-codex-spark", "gpt-5.5", "gpt-5.4-mini"],
   },
   /** Kimi's coding endpoint reached with a subscription token instead of a
    *  key — Kimi bills the two separately. */
@@ -255,7 +358,7 @@ export const PROVIDER_PRESETS = {
     baseUrl: "https://api.githubcopilot.com",
     auth: "oauth",
     defaultModel: "gpt-5.6-sol",
-    models: ["gpt-5.6-sol", "claude-opus-5", "claude-opus-4.8", "kimi-k2.7-code"],
+    models: ["gpt-5.6-sol", "claude-opus-5", "claude-sonnet-5", "gpt-6-astra"],
   },
 } as const satisfies Record<string, ProviderPreset>;
 

package/src/providers/factory.ts

@@ -91,6 +91,7 @@ export function createPresetProvider(id: ProviderPresetId, config: PresetProvide
         effort: config.effort,
         maxTokens: config.maxTokens,
         baseUrl: preset.baseUrl,
+        ...(preset.path ? { path: preset.path } : {}),
         headers,
         ...(config.providerOrder ? { providerOrder: config.providerOrder } : {}),
         fetchImpl: config.fetchImpl,

package/src/providers/openai.ts

@@ -402,8 +402,12 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
         request.provider = { order: config.providerOrder, allow_fallbacks: true };
       }
 
+      const defaultPath = /\/v\d+[^/]*$/i.test(baseUrl.replace(/\/+$/, ""))
+        ? "/chat/completions"
+        : "/v1/chat/completions";
+
       for await (const data of streamSse({
-        url: apiUrl(baseUrl, config.path ?? "/v1/chat/completions"),
+        url: apiUrl(baseUrl, config.path ?? defaultPath),
         headers: { authorization: `Bearer ${config.apiKey}`, ...config.headers },
         body: request,
         provider: id,

package/src/rate-limit.ts

@@ -39,6 +39,7 @@ export interface RateLimitReset {
 }
 
 const RETRY_AFTER = "retry-after";
+const RETRY_AFTER_MS = "retry-after-ms";
 const UNIFIED_5H = "anthropic-ratelimit-unified-5h-";
 const UNIFIED_7D = "anthropic-ratelimit-unified-7d-";
 
@@ -74,27 +75,50 @@ function unixSecondsToMs(value: string | null): number | undefined {
   return n === undefined ? undefined : n * 1000;
 }
 
+/** Matches Go-style duration strings like "2m59.56s", "7.66s", "500ms", "1h30m". */
+const DURATION_PATTERN =
+  /^(?:(\d+(?:\.\d+)?)h)?(?:(\d+(?:\.\d+)?)m(?!s))?(?:(\d+(?:\.\d+)?)m?s)?$/i;
+
+function parseDurationMs(text: string): number | undefined {
+  const match = DURATION_PATTERN.exec(text.trim());
+  if (!match || !match.slice(1).some(Boolean)) return undefined;
+  const hours = Number(match[1]) || 0;
+  const minutes = Number(match[2]) || 0;
+  const seconds = Number(match[3]) || 0;
+  const isMs = text.toLowerCase().endsWith("ms");
+  const secMs = isMs ? seconds : seconds * 1000;
+  const total = Math.round(hours * 3_600_000 + minutes * 60_000 + secMs);
+  return total >= 0 ? total : undefined;
+}
+
 /** Absolute reset time from a header that spells it either way: epoch seconds
- *  (the unified reset, most gateways) or an RFC 3339 / HTTP-date string (the
- *  API-key reset headers). The digit test runs first because `Date.parse("120")`
- *  is not a rejection — it is the year 120, two millennia in the past. */
+ *  (the unified reset, most gateways), Go-style duration ("2m59s" from OpenAI),
+ *  or an RFC 3339 / HTTP-date string (the API-key reset headers). */
 function resetHeaderMs(value: string | null, now: number): number | undefined {
   if (!value) return undefined;
-  if (/^\d+$/.test(value)) {
-    const seconds = Number(value);
+  const trimmed = value.trim();
+  if (/^\d+$/.test(trimmed)) {
+    const seconds = Number(trimmed);
     return seconds > EPOCH_SECONDS_FLOOR ? seconds * 1000 : now + seconds * 1000;
   }
-  const at = Date.parse(value);
+  const duration = parseDurationMs(trimmed);
+  if (duration !== undefined) {
+    return now + duration;
+  }
+  const at = Date.parse(trimmed);
   return Number.isNaN(at) ? undefined : at;
 }
 
-/** `retry-after` is delay-seconds, but RFC 9110 also allows an absolute date,
- *  and both forms occur across the providers this package speaks to. */
+/** `retry-after` is delay-seconds, Go-style duration ("60s"), or an absolute
+ *  HTTP date, and all three occur across providers this package speaks to. */
 function parseRetryAfter(value: string | null, now: number): number | undefined {
   if (!value) return undefined;
-  const n = Number(value);
+  const trimmed = value.trim();
+  const n = Number(trimmed);
   if (Number.isFinite(n) && n >= 0) return n * 1000;
-  const at = Date.parse(value);
+  const duration = parseDurationMs(trimmed);
+  if (duration !== undefined) return duration;
+  const at = Date.parse(trimmed);
   return Number.isNaN(at) ? undefined : Math.max(0, at - now);
 }
 
@@ -110,7 +134,10 @@ function parseRetryAfter(value: string | null, now: number): number | undefined
 export function parseRateLimitReset(headers: Headers, now = Date.now()): RateLimitReset {
   const result: RateLimitReset = {};
 
-  const after = parseRetryAfter(headers.get(RETRY_AFTER), now);
+  // `retry-after-ms` is Anthropic's millisecond spelling of the same header;
+  // when both ride a response the coarser one is the same deadline rounded.
+  const afterMs = numeric(headers.get(RETRY_AFTER_MS));
+  const after = afterMs ?? parseRetryAfter(headers.get(RETRY_AFTER), now);
   if (after !== undefined) {
     result.retryAfterMs = after;
     result.resetAtMs = now + after;
@@ -176,9 +203,31 @@ function jsonSeconds(value: unknown): number | undefined {
   return Number.isFinite(n) ? n : undefined;
 }
 
+const PROSE_RETRY_PATTERNS = [
+  /try again in (\d+(?:\.\d+)?)\s*(s(?:ec(?:ond)?s?)?|m(?:in(?:ute)?s?)?|h(?:(?:ou)?rs?)?)/i,
+  /retry after (\d+(?:\.\d+)?)\s*(s(?:ec(?:ond)?s?)?|m(?:in(?:ute)?s?)?|h(?:(?:ou)?rs?)?)/i,
+  /resets? (?:in|after) (\d+(?:\.\d+)?)\s*(s(?:ec(?:ond)?s?)?|m(?:in(?:ute)?s?)?|h(?:(?:ou)?rs?)?)/i,
+  /retry[- ]after[:\s]+(\d+(?:\.\d+)?)\s*(s(?:ec(?:ond)?s?)?|m(?:in(?:ute)?s?)?|h(?:(?:ou)?rs?)?)?/i,
+];
+
+function parseProseRetryMs(text: string): number | undefined {
+  for (const pattern of PROSE_RETRY_PATTERNS) {
+    const match = text.match(pattern);
+    if (!match?.[1]) continue;
+    const val = Number.parseFloat(match[1]);
+    if (!Number.isFinite(val) || val <= 0) continue;
+    const unit = (match[2] ?? "s").toLowerCase();
+    const multiplier = unit.startsWith("h") ? 3_600_000 : unit.startsWith("m") ? 60_000 : 1_000;
+    return Math.ceil(val * multiplier);
+  }
+  return undefined;
+}
+
 /**
  * ChatGPT's codex backend puts the reset IN THE 429 BODY, not the headers:
  * `{"error":{"type":"usage_limit_reached","resets_at":1788801754,"resets_in_seconds":2501465}}`.
+ * Other providers and gateways (LiteLLM, Cloudflare, Ollama) report countdowns in prose:
+ * `Rate limit exceeded. Try again in 20s.`
  * The window name is inferred from the wait itself, and only past the floor
  * above — a sub-minute retry is a throttle, not a subscription window.
  */
@@ -187,33 +236,46 @@ export function parseUsageLimitBody(bodyText: string, now = Date.now()): RateLim
   try {
     body = JSON.parse(bodyText);
   } catch {
-    return {};
+    body = undefined;
   }
-  if (typeof body !== "object" || body === null) return {};
-  // The fields sit under `error` on the documented shape and at the top level
-  // on some gateway relays of it. Both are read rather than guessed between.
-  const source = (body as Record<string, unknown>).error;
-  const error =
-    typeof source === "object" && source !== null
-      ? (source as Record<string, unknown>)
-      : (body as Record<string, unknown>);
-
-  const inSeconds = jsonSeconds(error.resets_in_seconds);
-  const atSeconds = jsonSeconds(error.resets_at);
-  const waitMs =
-    inSeconds !== undefined && inSeconds >= 0
-      ? inSeconds * 1000
-      : atSeconds !== undefined && atSeconds > 0
-        ? Math.max(0, atSeconds * 1000 - now)
-        : undefined;
-  if (waitMs === undefined) return {};
-
-  const result: RateLimitReset = { retryAfterMs: waitMs, resetAtMs: now + waitMs };
-  if (waitMs > WINDOW_FLOOR_MS) {
-    result.window =
-      waitMs <= FIVE_HOUR_MAX_MS ? "5h" : waitMs <= WEEKLY_MAX_MS ? "weekly" : "monthly";
+
+  if (typeof body === "object" && body !== null) {
+    const source = (body as Record<string, unknown>).error;
+    const error =
+      typeof source === "object" && source !== null
+        ? (source as Record<string, unknown>)
+        : (body as Record<string, unknown>);
+
+    const inSeconds = jsonSeconds(error.resets_in_seconds);
+    const atSeconds = jsonSeconds(error.resets_at);
+    const waitMs =
+      inSeconds !== undefined && inSeconds >= 0
+        ? inSeconds * 1000
+        : atSeconds !== undefined && atSeconds > 0
+          ? Math.max(0, atSeconds * 1000 - now)
+          : undefined;
+    if (waitMs !== undefined) {
+      const result: RateLimitReset = { retryAfterMs: waitMs, resetAtMs: now + waitMs };
+      if (waitMs > WINDOW_FLOOR_MS) {
+        result.window =
+          waitMs <= FIVE_HOUR_MAX_MS ? "5h" : waitMs <= WEEKLY_MAX_MS ? "weekly" : "monthly";
+      }
+      return result;
+    }
   }
-  return result;
+
+  // If structured JSON did not yield a reset, check prose wording
+  const waitMs = parseProseRetryMs(bodyText);
+  if (waitMs !== undefined) {
+    const result: RateLimitReset = { retryAfterMs: waitMs, resetAtMs: now + waitMs };
+    if (waitMs > WINDOW_FLOOR_MS) {
+      result.window =
+        waitMs <= FIVE_HOUR_MAX_MS ? "5h" : waitMs <= WEEKLY_MAX_MS ? "weekly" : "monthly";
+    }
+    return result;
+  }
+
+  return {};
 }
 
 /** Read both places an endpoint can report its reset. A short Retry-After is