Add sync for tensorx

.github/workflows/sync-models.yml

@@ -90,6 +90,7 @@ jobs:
           XAI_API_KEY: ${{ secrets.XAI_API_KEY }}
           CLOUDFLARE_WORKERS_AI_SYNC_ACCOUNT_ID: ${{ secrets.CLOUDFLARE_WORKERS_AI_SYNC_ACCOUNT_ID }}
           CLOUDFLARE_WORKERS_AI_SYNC_API_TOKEN: ${{ secrets.CLOUDFLARE_WORKERS_AI_SYNC_API_TOKEN }}
+          TENSORX_API_KEY: ${{ secrets.TENSORX_API_KEY }}
 
       - name: Validate models
         run: bun validate

packages/core/src/sync/index.ts

@@ -27,6 +27,7 @@ import { openrouter } from "./providers/openrouter.js";
 import { ovhcloud } from "./providers/ovhcloud.js";
 import { pioneer } from "./providers/pioneer.js";
 import { requesty } from "./providers/requesty.js";
+import { tensorx } from "./providers/tensorx.js";
 import { tinfoil } from "./providers/tinfoil.js";
 import { vercel } from "./providers/vercel.js";
 import { venice } from "./providers/venice.js";
@@ -133,6 +134,7 @@ export const providers: {
   ovhcloud: SyncProvider<any>;
   pioneer: SyncProvider<any>;
   requesty: SyncProvider<any>;
+  tensorx: SyncProvider<any>;
   tinfoil: SyncProvider<any>;
   vercel: SyncProvider<any>;
   venice: SyncProvider<any>;
@@ -161,6 +163,7 @@ export const providers: {
   ovhcloud,
   pioneer,
   requesty,
+  tensorx,
   tinfoil,
   vercel,
   venice,
@@ -183,7 +186,7 @@ export const groups = {
     "vercel",
   ],
   cloudflare: ["cloudflare-workers-ai"],
-  direct: ["ambient", "anthropic", "baseten", "chutes", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
+  direct: ["ambient", "anthropic", "baseten", "chutes", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tensorx", "tinfoil", "venice", "wandb", "xai"],
 } as const;
 
 type ProviderID = keyof typeof providers;

packages/core/src/sync/providers/tensorx.ts

@@ -0,0 +1,208 @@
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://api.tensorx.ai/v1/model/info";
+
+// `supported_openai_params` is a static LiteLLM-style list: every catalog entry
+// advertises the same params, including `temperature` and `response_format` on
+// the Whisper and embedding models. It carries no per-model signal, so nothing
+// is derived from it — `temperature` and `structured_output` stay lab metadata.
+//
+// `supports_*` uses three values: true, false, and null for "not published".
+// Only real booleans are authored; null leaves the field to the base model.
+const TensorXModelInfo = z.object({
+  mode: z.string().nullish(),
+  max_input_tokens: z.number().int().nonnegative().nullish(),
+  max_output_tokens: z.number().int().nonnegative().nullish(),
+  max_tokens: z.number().int().nonnegative().nullish(),
+  supports_reasoning: z.boolean().nullish(),
+  supports_tool_choice: z.boolean().nullish(),
+  supports_function_calling: z.boolean().nullish(),
+  supports_vision: z.boolean().nullish(),
+  input_cost_per_token: z.number().nonnegative().nullish(),
+  output_cost_per_token: z.number().nonnegative().nullish(),
+  cache_read_input_token_cost: z.number().nonnegative().nullish(),
+  cache_creation_input_token_cost: z.number().nonnegative().nullish(),
+}).passthrough();
+
+export const TensorXModel = z.object({
+  model_name: z.string().min(1),
+  model_info: TensorXModelInfo,
+});
+
+export const TensorXResponse = z.object({
+  data: z.array(TensorXModel),
+});
+
+export type TensorXModel = z.infer<typeof TensorXModel>;
+
+export const tensorx = {
+  id: "tensorx",
+  name: "TensorX",
+  modelsDir: "providers/tensorx/models",
+  // /v1/model/info carries no reasoning controls and no side-channel field, so
+  // a created reasoner would be published with the runner's fallback
+  // `reasoning_options = []` — an assertion of "no caller control" that nothing
+  // here backs — and no `interleaved`. New IDs are reported for hand-authoring
+  // instead; updates to existing TOMLs are unaffected.
+  skipCreates: true,
+  preserveBaseModels: false,
+  // /v1/model/info returns a per-key view: a key scoped to a model group sees
+  // only that group, and chat requests for the rest fail with 403 rather than
+  // 404. Absence from the response is therefore not evidence that a model was
+  // retired, so local entries are never deleted on the strength of it.
+  deleteMissing: false,
+  sourceID(model) {
+    // Non-chat entries (embeddings, transcription, speech) are out of scope and
+    // are dropped silently; only unauthorable chat models get reported.
+    return model.model_info.mode === "chat" ? model.model_name : undefined;
+  },
+  skippedNotice(ids) {
+    if (ids.length === 0) return [];
+    return [
+      `${ids.length} TensorX chat models were not created: /v1/model/info publishes no name, description, or release date, and no reasoning controls or side-channel field, so a complete model cannot be authored safely.`,
+      `Add the lab entry under \`models/\` and hand-author \`reasoning_options\` / \`interleaved\` against the live API. Skipped remote IDs: ${ids.map((id) => `\`${id}\``).join(", ")}`,
+    ];
+  },
+  missingNotice(paths) {
+    if (paths.length === 0) return [];
+    return [
+      `${paths.length} local TensorX models were absent from /v1/model/info and were retained for manual lifecycle review.`,
+      `That endpoint is a per-key view, so absence can mean the sync key lacks access rather than the model being retired: ${paths.map((path) => `\`${path}\``).join(", ")}`,
+    ];
+  },
+  async fetchModels() {
+    const apiKey = process.env.TENSORX_API_KEY;
+    if (!apiKey) {
+      throw new Error("TensorX sync requires TENSORX_API_KEY environment variable");
+    }
+    const response = await fetch(API_ENDPOINT, {
+      headers: { Authorization: `Bearer ${apiKey}` },
+    });
+    if (!response.ok) {
+      throw new Error(`TensorX model info request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    const models = TensorXResponse.parse(raw).data;
+    const seen = new Set<string>();
+    return models.filter((model) => {
+      if (seen.has(model.model_name)) return false;
+      seen.add(model.model_name);
+      return true;
+    });
+  },
+  translateModel(model, context) {
+    if (model.model_info.mode !== "chat") return undefined;
+
+    const existing = context.existing(model.model_name);
+    // A base_model already authored locally wins over re-derivation. With
+    // preserveBaseModels false the runner will not put it back, so a resolution
+    // miss here (dropped alias, renamed lab entry, ID drift) would otherwise
+    // flatten the inherited lab fields into the provider TOML.
+    const baseModel = existing?.base_model ?? resolveBaseModel(model.model_name);
+    // Nothing in this catalog can stand in for lab metadata — no display name,
+    // no description, no release date — so an unknown model is reported for
+    // hand-authoring instead of being invented.
+    if (baseModel === undefined && existing === undefined) return undefined;
+
+    const built = buildTensorXModel(model, baseModel, existing);
+    return built === undefined ? undefined : { id: model.model_name, model: built };
+  },
+} satisfies SyncProvider<TensorXModel>;
+
+function buildTensorXModel(
+  model: TensorXModel,
+  baseModel: string | undefined,
+  existing: ExistingModel | undefined,
+): SyncedModel | undefined {
+  const info = model.model_info;
+
+  const limit = {
+    context: info.max_input_tokens ?? info.max_tokens ?? existing?.limit?.context,
+    output: info.max_output_tokens ?? existing?.limit?.output,
+  };
+
+  const input = perMillion(info.input_cost_per_token) ?? existing?.cost?.input;
+  const output = perMillion(info.output_cost_per_token) ?? existing?.cost?.output;
+  // Pricing is published only as a complete pair. A base_model file validates
+  // against a deepPartial schema, so a half-resolved cost would be written out
+  // as real pricing rather than rejected; a full model would abort the sync.
+  const cost = input === undefined || output === undefined
+    ? existing?.cost
+    : {
+      ...existing?.cost,
+      input,
+      output,
+      cache_read: perMillion(info.cache_read_input_token_cost) ?? existing?.cost?.cache_read,
+      // cache_creation_input_token_cost is null for every model in the catalog,
+      // so a null is "not published" rather than "not charged". Keep the
+      // authored price; real values publish as soon as TensorX fills it in.
+      cache_write: perMillion(info.cache_creation_input_token_cost) ?? existing?.cost?.cache_write,
+    };
+
+  // Never bring a brand-new model into the catalog without real pricing.
+  if (existing === undefined && cost === undefined) return undefined;
+
+  const toolFlags = [info.supports_tool_choice, info.supports_function_calling]
+    .filter((flag) => flag !== null && flag !== undefined);
+
+  // supports_vision is the only modality signal the catalog carries, so
+  // `attachment` and `modalities.input` move together. Only `image` is edited —
+  // the flag says nothing about video or pdf — and `attachment` is then read
+  // back off the resulting list rather than off the flag, so a model that keeps
+  // video after losing image stays `attachment = true`.
+  const vision = info.supports_vision;
+  const inheritedInput = existing?.modalities?.input;
+  const modalities = vision === null || vision === undefined || inheritedInput === undefined
+    ? existing?.modalities
+    : {
+      ...existing?.modalities,
+      input: vision
+        ? (inheritedInput.includes("image") ? inheritedInput : [...inheritedInput, "image"])
+        : inheritedInput.filter((modality) => modality !== "image"),
+    };
+  const resolvedInput = modalities?.input;
+  const attachment = vision === null || vision === undefined || resolvedInput === undefined
+    ? existing?.attachment
+    : resolvedInput.some((modality) => modality !== "text");
+
+  // `existing` is the base-model-resolved view, so factorBaseModel drops every
+  // field that still matches the lab entry and keeps only the real deltas.
+  const values: Record<string, unknown> = {
+    ...existing,
+    attachment,
+    reasoning: info.supports_reasoning ?? existing?.reasoning,
+    tool_call: toolFlags.length > 0 ? toolFlags.some(Boolean) : existing?.tool_call,
+    cost,
+    limit,
+    modalities,
+  };
+  delete values.base_model;
+  delete values.base_model_omit;
+
+  return baseModel === undefined
+    ? values as SyncedFullModel
+    : factorBaseModel(baseModel, values, limit, existing?.base_model_omit);
+}
+
+// Dated snapshots TensorX serves that have no lab entry of their own.
+const BASE_MODEL_ALIASES: Record<string, string | undefined> = {
+  "deepseek/deepseek-r1-0528": "deepseek/deepseek-r1",
+  "deepseek/deepseek-r1-0625": "deepseek/deepseek-r1",
+};
+
+function resolveBaseModel(modelID: string): string | undefined {
+  // resolveCanonicalBaseModel owns the org-prefix map and matches metadata
+  // filenames case-insensitively, so `minimax/minimax-m3` still resolves to
+  // `models/minimax/MiniMax-M3.toml`.
+  return resolveCanonicalBaseModel(BASE_MODEL_ALIASES[modelID] ?? modelID);
+}
+
+function perMillion(costPerToken: number | null | undefined): number | undefined {
+  if (costPerToken === null || costPerToken === undefined) return undefined;
+  return Math.round(costPerToken * 1_000_000 * 1e10) / 1e10;
+}

providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml

@@ -1,3 +1,10 @@
+# Not reachable with the sync key (403, models=['public']), so nothing here
+# could be measured against this host. providers/openrouter/models/deepseek/
+# deepseek-chat-v3.1.toml authors a toggle, so that control shape is copied
+# rather than keeping a full effort enum.
+# Toggle: chat_template_kwargs.thinking = true|false — the path every reachable
+# DeepSeek model on this host uses, but unverified for this ID.
+# [interleaved] is omitted rather than assumed, for the same reason.
 name = "DeepSeek Chat V3.1"
 description = "DeepSeek chat model for instruction following, coding, and analysis"
 family = "deepseek"
@@ -10,10 +17,8 @@ tool_call = true
 knowledge = "2024-11"
 open_weights = true
 
-
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+type = "toggle"
 
 [cost]
 input = 0.2

providers/tensorx/models/deepseek/deepseek-r1-0528.toml

@@ -1,11 +1,14 @@
+# reasoning_effort is accepted but does not grade depth on this host, and none
+# is rejected (400), so reasoning is always on. Matches the OpenRouter entry for
+# this model.
 base_model = "deepseek/deepseek-r1"
 name = "DeepSeek R1-0528"
 release_date = "2025-05-28"
 last_updated = "2025-05-28"
+reasoning_options = []
 
-[[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; reasoning is mandatory, "none" is rejected
-values = ["minimal", "low", "medium", "high", "xhigh", "max"]
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 0.66

providers/tensorx/models/deepseek/deepseek-v3.2.toml

@@ -1,3 +1,8 @@
+# Effort: reasoning_effort = none|high. This host does not reason by default —
+# reasoning_content is empty unless an effort value is sent — and none keeps it
+# off. chat_template_kwargs.enable_thinking and thinking are no-ops. Measured
+# by reasoning_content length; invalid values return 400 and no graded level
+# was separable, so one on level is published.
 name = "DeepSeek V3.2"
 description = "DeepSeek chat model for instruction following, coding, and analysis"
 family = "deepseek"
@@ -10,10 +15,12 @@ tool_call = true
 knowledge = "2025-05"
 open_weights = true
 
-
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+type = "effort"
+values = ["none", "high"]
+
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 0.3

providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml

@@ -1,12 +1,21 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted but not honored. Measured by reasoning_content
+# length — not usage.reasoning_tokens, which this host under-reports — low,
+# high, max and an invalid value all fall in the same range, so this host is
+# toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-flash-0731"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 0.25
-output = 0.30
-cache_read = 0.06
+output = 0.3
+cache_read = 0.0625
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000

providers/tensorx/models/deepseek/deepseek-v4-flash.toml

@@ -1,7 +1,15 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted but not honored. Measured by reasoning_content
+# length — not usage.reasoning_tokens, which this host under-reports — low,
+# high, max and an invalid value all fall in the same range, so this host is
+# toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-flash"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 0.15
@@ -10,4 +18,5 @@ cache_read = 0.0375
 cache_write = 0.1875
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000

providers/tensorx/models/deepseek/deepseek-v4-pro.toml

@@ -1,7 +1,15 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted but not honored. Measured by reasoning_content
+# length — not usage.reasoning_tokens, which this host under-reports — low,
+# high, max and an invalid value all fall in the same range, so this host is
+# toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-pro"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 1.75
@@ -10,4 +18,5 @@ cache_read = 0.4375
 cache_write = 2.185
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000