@shop-os/modelkit 0.1.0 → 0.1.1

package/README.md

@@ -101,6 +101,11 @@ fetch, logging, and the `onUsage` sink.
 pass-through instances; catalog rows carry their own and the caller's are dropped
 with an `option_ignored` warning.
 
+Text catalog rows carry no reasoning defaults. A catalog id behaves like the
+provider's own default, so `temperature` and friends work as they do in `ai`.
+Opt into thinking per call, for example
+`providerOptions: { anthropic: { thinking: { type: "enabled", budgetTokens: 4000 } } }`.
+
 The kit keeps its `Warning` type for coercions; the SDK's warning type is `AiWarning`.
 `languageModel()` remains the low-level handle for Mastra and custom middleware.
 

package/dist/index.cjs

@@ -212,6 +212,44 @@ var IMAGE_MODELS = [
     usage: { ...IMAGE_USAGE },
     pricingKind: "image_size"
   },
+  {
+    // Same Images API surface as gpt-image-2 plus the "xhigh" and "max" quality tiers.
+    id: "gpt-image-2.5-sunburst",
+    aliases: [],
+    label: "GPT Image 2.5 Sunburst",
+    type: "image",
+    adapter: "openai-image",
+    providerModelId: "gpt-image-2.5-sunburst",
+    lifecycle: "active",
+    constraints: {
+      aspectRatios: ["1:1", "3:2", "2:3", "16:9", "9:16", "4:5", "5:4", "adaptive"],
+      sizes: ["1K", "2K", "4K"],
+      references: { images: 16 },
+      supports: ["quality", "count"]
+    },
+    defaults: { size: "1K", aspectRatio: "adaptive", quality: "auto" },
+    usage: { ...IMAGE_USAGE },
+    pricingKind: "image_size"
+  },
+  {
+    // Same Images API surface as gpt-image-2 plus the "xhigh" and "max" quality tiers.
+    id: "gpt-image-2.5-flare",
+    aliases: [],
+    label: "GPT Image 2.5 Flare",
+    type: "image",
+    adapter: "openai-image",
+    providerModelId: "gpt-image-2.5-flare",
+    lifecycle: "active",
+    constraints: {
+      aspectRatios: ["1:1", "3:2", "2:3", "16:9", "9:16", "4:5", "5:4", "adaptive"],
+      sizes: ["1K", "2K", "4K"],
+      references: { images: 16 },
+      supports: ["quality", "count"]
+    },
+    defaults: { size: "1K", aspectRatio: "adaptive", quality: "auto" },
+    usage: { ...IMAGE_USAGE },
+    pricingKind: "image_size"
+  },
   {
     // Cowork allows 7 ratios; trigger-workflows allows 9 (adds 3:2, 9:21, drops 4:5/5:4).
     id: "muse-image-1.0",
@@ -588,77 +626,42 @@ function text(id, label, providerModelId, extra) {
     ...extra
   };
 }
-var ANTHROPIC_THINKING = {
-  anthropic: {
-    thinking: { type: "enabled", budgetTokens: 4e3 },
-    sendReasoning: true
-  }
-};
-var ANTHROPIC_ADAPTIVE = {
-  anthropic: {
-    thinking: { type: "adaptive", display: "summarized" },
-    sendReasoning: true
-  }
-};
-var OPENAI_REASONING = {
-  openai: {
-    reasoningEffort: "medium",
-    reasoningSummary: "auto"
-  }
-};
-var DEEPSEEK_THINKING = {
-  deepseek: {
-    thinking: { type: "enabled" }
-  }
-};
 var TEXT_MODELS = [
   text("gpt-4o", "GPT-4o", "gpt-4o"),
   text("gpt-4o-mini", "GPT-4o mini", "gpt-4o-mini"),
   text("gpt-5.1", "GPT-5.1", "gpt-5.1"),
-  text("gpt-5.1-mini", "GPT-5.1 mini", "gpt-5.1-mini"),
   text("gpt-5.2", "GPT-5.2", "gpt-5.2"),
-  text("gpt-5.2-mini", "GPT-5.2 mini", "gpt-5.2-mini"),
   text("gpt-5.4", "GPT-5.4", "gpt-5.4"),
-  text("gpt-5.4-mini", "GPT-5.4 mini", "gpt-5.4-mini"),
-  text("gpt-5.5", "GPT-5.5", "gpt-5.5"),
-  text("gpt-5.6-luna", "GPT-5.6 Luna", "gpt-5.6-luna", {
-    providerOptions: { ...OPENAI_REASONING }
-  }),
-  text("gpt-5.6-terra", "GPT-5.6 Terra", "gpt-5.6-terra", {
-    providerOptions: { ...OPENAI_REASONING }
-  }),
-  text("gpt-5.6-sol", "GPT-5.6 Sol", "gpt-5.6-sol", {
-    providerOptions: { ...OPENAI_REASONING }
-  }),
-  // Dated 4.0 ids are what vibe-api, pro-admin and the TW geo agents still send.
-  text("claude-sonnet-4", "Claude Sonnet 4", "claude-sonnet-4-20250514", {
-    aliases: ["claude-sonnet-4.0", "claude-sonnet-4-20250514"]
-  }),
-  text("claude-opus-4", "Claude Opus 4", "claude-opus-4-20250514", {
-    aliases: ["claude-opus-4.0", "claude-opus-4-20250514"]
+  // OpenAI never shipped gpt-5.1-mini or gpt-5.2-mini; callers that still send them get 5.4 mini.
+  text("gpt-5.4-mini", "GPT-5.4 mini", "gpt-5.4-mini", {
+    aliases: ["gpt-5.1-mini", "gpt-5.2-mini"]
   }),
+  text("gpt-5.5", "GPT-5.5", "gpt-5.5"),
+  text("gpt-5.6-luna", "GPT-5.6 Luna", "gpt-5.6-luna"),
+  text("gpt-5.6-terra", "GPT-5.6 Terra", "gpt-5.6-terra"),
+  text("gpt-5.6-sol", "GPT-5.6 Sol", "gpt-5.6-sol"),
   // Library key "claude-sonnet-4.5" maps to the SDK id "claude-sonnet-4-5".
   text("claude-sonnet-4-5", "Claude Sonnet 4.5", "claude-sonnet-4-5", {
     aliases: ["claude-sonnet-4.5"]
   }),
+  // Anthropic retired the dated 4.0 ids; callers that still send them get 4.6.
   text("claude-sonnet-4-6", "Claude Sonnet 4.6", "claude-sonnet-4-6", {
-    aliases: ["claude-sonnet-4.6"],
-    providerOptions: { ...ANTHROPIC_THINKING }
-  }),
-  text("claude-sonnet-5", "Claude Sonnet 5", "claude-sonnet-5", {
-    providerOptions: { ...ANTHROPIC_ADAPTIVE }
+    aliases: [
+      "claude-sonnet-4.6",
+      "claude-sonnet-4",
+      "claude-sonnet-4.0",
+      "claude-sonnet-4-20250514"
+    ]
   }),
+  text("claude-sonnet-5", "Claude Sonnet 5", "claude-sonnet-5"),
   text("claude-opus-4-8", "Claude Opus 4.8", "claude-opus-4-8", {
-    aliases: ["claude-opus-4.8"],
-    providerOptions: { ...ANTHROPIC_ADAPTIVE }
+    aliases: ["claude-opus-4.8", "claude-opus-4", "claude-opus-4.0", "claude-opus-4-20250514"]
   }),
   text("claude-haiku-4-5", "Claude Haiku 4.5", "claude-haiku-4-5", {
-    aliases: ["claude-haiku-4.5", "claude-haiku-4-5-20251001"],
-    providerOptions: { ...ANTHROPIC_THINKING }
+    aliases: ["claude-haiku-4.5", "claude-haiku-4-5-20251001"]
   }),
   text("claude-opus-5", "Claude Opus 5", "claude-opus-5", {
-    aliases: ["claude-opus-5.0"],
-    providerOptions: { ...ANTHROPIC_ADAPTIVE }
+    aliases: ["claude-opus-5.0"]
   }),
   text("claude-fable-5", "Claude Fable 5", "claude-fable-5"),
   text("gemini-2.5-flash", "Gemini 2.5 Flash", "gemini-2.5-flash"),
@@ -677,12 +680,8 @@ var TEXT_MODELS = [
   text("sonar", "Sonar", "sonar"),
   text("sonar-pro", "Sonar Pro", "sonar-pro"),
   text("sonar-reasoning-pro", "Sonar Reasoning Pro", "sonar-reasoning-pro"),
-  text("deepseek-v4-flash", "DeepSeek V4 Flash", "deepseek-v4-flash", {
-    providerOptions: { ...DEEPSEEK_THINKING }
-  }),
-  text("deepseek-v4-pro", "DeepSeek V4 Pro", "deepseek-v4-pro", {
-    providerOptions: { ...DEEPSEEK_THINKING }
-  }),
+  text("deepseek-v4-flash", "DeepSeek V4 Flash", "deepseek-v4-flash"),
+  text("deepseek-v4-pro", "DeepSeek V4 Pro", "deepseek-v4-pro"),
   text("grok-4.3", "Grok 4.3", "grok-4.3"),
   text("kimi-k2.5", "Kimi K2.5", "kimi-k2.5"),
   text("kimi-k3", "Kimi K3", "kimi-k3"),
@@ -1513,23 +1512,46 @@ async function withRetry(fn, opts = {}) {
 }
 function sleep(ms, signal) {
   return new Promise((resolve, reject) => {
-    const t = setTimeout(resolve, ms);
-    signal?.addEventListener(
-      "abort",
-      () => {
-        clearTimeout(t);
-        reject(
-          new ModelKitError({ code: "timeout", message: "aborted", retryable: false })
-        );
-      },
-      { once: true }
-    );
+    const rejectAborted = () => {
+      reject(new ModelKitError({ code: "timeout", message: "aborted", retryable: false }));
+    };
+    if (signal?.aborted) {
+      rejectAborted();
+      return;
+    }
+    let timer;
+    const onAbort = () => {
+      if (timer !== void 0) clearTimeout(timer);
+      rejectAborted();
+    };
+    timer = setTimeout(() => {
+      signal?.removeEventListener("abort", onAbort);
+      resolve();
+    }, ms);
+    signal?.addEventListener("abort", onAbort, { once: true });
   });
 }
+function throwAborted() {
+  return new ModelKitError({ code: "timeout", message: "aborted", retryable: false });
+}
+function transportError(provider, cause, model) {
+  const message = cause instanceof Error ? cause.message : String(cause);
+  const aborted = cause instanceof Error && cause.name === "AbortError";
+  const init = {
+    code: aborted ? "timeout" : "network",
+    message: `${provider} request failed before a response: ${message}`,
+    provider,
+    retryable: !aborted,
+    cause
+  };
+  if (model !== void 0) init.model = model;
+  return new ModelKitError(init);
+}
 async function pollUntil(check, opts) {
   const start = Date.now();
   let interval = opts.intervalMs;
   for (; ; ) {
+    if (opts.signal?.aborted) throw throwAborted();
     const value = await check();
     if (value !== void 0) return value;
     if (Date.now() - start > opts.timeoutMs)
@@ -1538,6 +1560,7 @@ async function pollUntil(check, opts) {
         message: `poll timed out after ${opts.timeoutMs}ms`,
         retryable: false
       });
+    if (opts.signal?.aborted) throw throwAborted();
     await sleep(interval, opts.signal);
     if (opts.backoff)
       interval = Math.min(opts.maxIntervalMs ?? interval * 4, interval * opts.backoff);
@@ -1715,7 +1738,7 @@ var OPENAI_COMPATIBLE_BASE_URLS = {
 var GATEWAY_SLUGS = {
   openai: "openai",
   anthropic: "anthropic",
-  google: "google-ai-studio",
+  google: "google-ai-studio/v1beta",
   xai: "grok",
   perplexity: "perplexity-ai",
   deepseek: "deepseek"
@@ -2141,13 +2164,21 @@ function createModelKit(config, adapters) {
         });
       const job = await adapter.generateVideo(request, ctx);
       const result = job.result.bind(job);
+      let settled;
       return {
         ...job,
         warnings,
-        async result(opts) {
-          const out = await result(opts);
-          await emit({ ...out.usage, model: model.id });
-          return out;
+        result(opts) {
+          if (settled) return settled;
+          const pending = result(opts).then(
+            (out) => emit({ ...out.usage, model: model.id }).then(() => out),
+            (cause) => {
+              settled = void 0;
+              throw cause;
+            }
+          );
+          settled = pending;
+          return pending;
         }
       };
     },
@@ -2961,7 +2992,7 @@ async function generateFalVideo(req, ctx, deps) {
   if (isKling(providerModelId)) {
     const input = klingO1Input(req);
     if (ctx.routing.fal === "cloudflare-passthrough") {
-      return passthroughVideoJob(ctx, req, KLING_O1_ENDPOINT, input);
+      return await passthroughVideoJob(ctx, req, KLING_O1_ENDPOINT, input);
     }
     const result2 = await runJob(ctx, deps, KLING_O1_ENDPOINT, input);
     return completedVideoJob(
@@ -2996,7 +3027,7 @@ async function generateFalVideo(req, ctx, deps) {
     base.aspect_ratio = req.aspectRatio ?? "adaptive";
   }
   if (ctx.routing.fal === "cloudflare-passthrough") {
-    return passthroughVideoJob(ctx, req, endpoint, base);
+    return await passthroughVideoJob(ctx, req, endpoint, base);
   }
   const result = await runJob(ctx, deps, endpoint, base);
   return completedVideoJob(ctx, req, extractVideoUrl(result, endpoint, ctx.model.id), result);
@@ -3014,31 +3045,26 @@ function completedVideoJob(ctx, req, url, raw) {
     }
   };
 }
-function passthroughVideoJob(ctx, req, endpoint, input) {
+async function passthroughVideoJob(ctx, req, endpoint, input) {
   const cfg = gatewayConfig(ctx);
-  let statusUrl;
-  const submit = (async () => {
-    const res = await gatewayFetch(cfg, ctx.fetch, `${FAL_QUEUE_HOST}/${endpoint}`, {
-      method: "POST",
-      body: JSON.stringify({ ...input, sync_mode: true })
+  const res = await gatewayFetch(cfg, ctx.fetch, `${FAL_QUEUE_HOST}/${endpoint}`, {
+    method: "POST",
+    body: JSON.stringify({ ...input, sync_mode: true })
+  });
+  if (!res.ok) throw falError(res.status, await readBody(res), ctx.model.id);
+  const parsed = falQueueSubmitSchema.safeParse(await res.json());
+  if (!parsed.success) {
+    throw new ModelKitError({
+      code: "provider_rejected",
+      message: `fal ${endpoint} submit returned no status URL`,
+      provider: "fal",
+      model: ctx.model.id,
+      retryable: false
     });
-    if (!res.ok) throw falError(res.status, await readBody(res), ctx.model.id);
-    const parsed = falQueueSubmitSchema.safeParse(await res.json());
-    if (!parsed.success) {
-      throw new ModelKitError({
-        code: "provider_rejected",
-        message: `fal ${endpoint} submit returned no status URL`,
-        provider: "fal",
-        model: ctx.model.id,
-        retryable: false
-      });
-    }
-    statusUrl = parsed.data.status_url;
-    return parsed.data;
-  })();
+  }
+  const { request_id: jobId, status_url, response_url } = parsed.data;
   const checkStatus = async () => {
-    const { response_url } = await submit;
-    const statusRes = await gatewayFetch(cfg, ctx.fetch, statusUrl ?? "", { method: "GET" });
+    const statusRes = await gatewayFetch(cfg, ctx.fetch, status_url, { method: "GET" });
     if (!statusRes.ok)
       throw falError(statusRes.status, await readBody(statusRes), ctx.model.id);
     const statusParsed = falQueueStatusSchema.safeParse(await statusRes.json());
@@ -3052,7 +3078,7 @@ function passthroughVideoJob(ctx, req, endpoint, input) {
   };
   const duration = req.durationSeconds ?? 5;
   return {
-    jobId: endpoint,
+    jobId,
     async poll() {
       try {
         const url = await checkStatus();
@@ -4618,6 +4644,7 @@ var openaiImageAdapter = {
     const refs = req.references ?? [];
     const isEdit = refs.length > 0;
     const quality = req.quality ?? (isEdit ? "medium" : "auto");
+    const sdkQuality = quality;
     const count = req.count !== void 0 && req.count > 1 ? { n: req.count } : {};
     const retryOptions = {};
     if (ctx.signal !== void 0) retryOptions.signal = ctx.signal;
@@ -4629,14 +4656,14 @@ var openaiImageAdapter = {
               model: ctx.model.providerModelId,
               prompt: req.prompt,
               size,
-              quality,
+              quality: sdkQuality,
               ...count
             });
           }
           return await client2.images.generate({
             model: ctx.model.providerModelId,
             prompt: req.prompt,
-            quality,
+            quality: sdkQuality,
             ...count
           });
         }
@@ -4655,7 +4682,7 @@ var openaiImageAdapter = {
             image: imageInput,
             prompt: req.prompt,
             size,
-            quality,
+            quality: sdkQuality,
             ...count
           });
         }
@@ -4663,7 +4690,7 @@ var openaiImageAdapter = {
           model: ctx.model.providerModelId,
           image: imageInput,
           prompt: req.prompt,
-          quality,
+          quality: sdkQuality,
           ...count
         });
       } catch (cause) {
@@ -4884,7 +4911,7 @@ async function generateProdiaImage(req, ctx) {
         });
       } catch (err) {
         if (isModelKitError(err)) throw err;
-        throw prodiaError(void 0, String(err), ctx.model.id);
+        throw transportError("prodia", err, ctx.model.id);
       }
       if (!attemptRes.ok) {
         const text2 = await attemptRes.text().catch(() => "");
@@ -4894,7 +4921,7 @@ async function generateProdiaImage(req, ctx) {
     }, retryOpts);
   } catch (err) {
     if (isModelKitError(err)) throw err;
-    throw prodiaError(void 0, String(err), ctx.model.id);
+    throw transportError("prodia", err, ctx.model.id);
   }
   const bytes = parseImageResponse(
     new Uint8Array(await res.arrayBuffer()),
@@ -5182,6 +5209,27 @@ function klingInput(req) {
   if (req.negativePrompt !== void 0) input.negative_prompt = req.negativePrompt;
   return input;
 }
+function klingOmniInput(req) {
+  const images = (req.firstFrame ? [req.firstFrame] : []).map(refToUrl3);
+  if (req.lastFrame) images.push(refToUrl3(req.lastFrame));
+  const input = {
+    prompt: req.prompt,
+    duration: req.durationSeconds ?? 10,
+    mode: req.mode ?? "standard",
+    generate_audio: req.generateAudio ?? false
+  };
+  const startImage = images[0];
+  if (startImage !== void 0) {
+    input.start_image = startImage;
+    const endImage = images[1];
+    if (endImage !== void 0) input.end_image = endImage;
+  } else {
+    input.aspect_ratio = req.aspectRatio ?? "16:9";
+  }
+  const refs = (req.referenceImages ?? []).map(refToUrl3);
+  if (refs.length > 0) input.reference_images = refs;
+  return input;
+}
 function hailuoInput(req) {
   const input = {
     prompt: req.prompt,
@@ -5204,7 +5252,7 @@ function wanInput(req) {
 }
 async function generateReplicateVideo(req, ctx, deps) {
   const id = isLower(ctx.model.providerModelId);
-  const input = id.includes("kling") ? klingInput(req) : id.includes("hailuo") ? hailuoInput(req) : wanInput(req);
+  const input = id.includes("kling") ? id.includes("omni") ? klingOmniInput(req) : klingInput(req) : id.includes("hailuo") ? hailuoInput(req) : wanInput(req);
   const api = client(ctx, deps);
   const createOptions = { ...route(ctx.model.providerModelId), input };
   let prediction;
@@ -5294,6 +5342,9 @@ function lyriaAudioInput(req, ctx) {
   if (negative) input.negative_prompt = negative;
   return input;
 }
+function elevenLabsMusicDuration(seconds) {
+  return Math.min(300, Math.max(5, seconds ?? 15));
+}
 function elevenLabsMusicInput(req, ctx) {
   const prompt = req.prompt.trim();
   if (!prompt) {
@@ -5305,10 +5356,9 @@ function elevenLabsMusicInput(req, ctx) {
       retryable: false
     });
   }
-  const seconds = Math.min(300, Math.max(5, req.durationSeconds ?? 15));
   return {
     prompt,
-    music_length_ms: seconds * 1e3,
+    music_length_ms: elevenLabsMusicDuration(req.durationSeconds) * 1e3,
     force_instrumental: req.instrumental ?? true,
     output_format: "wav_cd_quality"
   };
@@ -5348,7 +5398,7 @@ async function generateReplicateAudio(req, ctx, deps) {
   const input = id.includes("mmaudio") ? mmaudioInput(req, ctx) : id.includes("elevenlabs") ? elevenLabsMusicInput(req, ctx) : lyriaAudioInput(req, ctx);
   const { url, predictTime, raw } = await runPrediction(ctx, input, deps);
   const usage2 = audioUsage(ctx, req);
-  usage2.quantity = ctx.model.usage.unit === "seconds" ? req.durationSeconds ?? 0 : predictTime;
+  usage2.quantity = ctx.model.usage.unit === "seconds" ? elevenLabsMusicDuration(req.durationSeconds) : predictTime;
   return { audio: { url }, usage: usage2, raw: predictionRaw(raw) };
 }