From dd95c9cf328cb0e2658b2b57e9e95790902a5183 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 06:46:42 +0200
Subject: [PATCH 01/14] feat(tensorx): add sync module for automated catalog
 updates
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Sync provider models from TensorX's own /v1/model/info endpoint.

Capability flags are three-valued in that catalog: true, false, and null
for "not published". Only real booleans are authored so a null leaves the
field inherited from the lab entry rather than overriding it with false.

supported_openai_params is deliberately not read. It is a static
LiteLLM-style list — all 32 catalog entries advertise the same params,
including temperature and response_format on the Whisper and embedding
models — so it carries no per-model signal. temperature and
structured_output stay lab metadata.

The endpoint publishes no name, description, or release date, so an
unknown model is reported through skippedNotice for hand-authoring
instead of being created from invented metadata.

Base models resolve through the shared resolveCanonicalBaseModel, which
matches metadata filenames case-insensitively.
---
 .github/workflows/sync-models.yml           |   1 +
 packages/core/src/sync/index.ts             |   5 +-
 packages/core/src/sync/providers/tensorx.ts | 162 ++++++++++++++++++++
 3 files changed, 167 insertions(+), 1 deletion(-)
 create mode 100644 packages/core/src/sync/providers/tensorx.ts

diff --git a/.github/workflows/sync-models.yml b/.github/workflows/sync-models.yml
index 7edd032424..ebb4f04304 100644
--- a/.github/workflows/sync-models.yml
+++ b/.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
diff --git a/packages/core/src/sync/index.ts b/packages/core/src/sync/index.ts
index a3fa90270f..69069ae457 100644
--- a/packages/core/src/sync/index.ts
+++ b/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;
diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
new file mode 100644
index 0000000000..ab0e02f61c
--- /dev/null
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -0,0 +1,162 @@
+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",
+  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 because /v1/model/info publishes no name, description, or release date to author a model from.`,
+      `Add the lab entry under \`models/\` (or hand-author the provider file) first. Skipped remote IDs: ${ids.map((id) => `\`${id}\``).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);
+    const baseModel = 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;
+
+    return {
+      id: model.model_name,
+      model: buildTensorXModel(model, baseModel, existing),
+    };
+  },
+} satisfies SyncProvider<TensorXModel>;
+
+function buildTensorXModel(
+  model: TensorXModel,
+  baseModel: string | undefined,
+  existing: ExistingModel | undefined,
+): SyncedModel {
+  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 cost = {
+    ...existing?.cost,
+    input: perMillion(info.input_cost_per_token) ?? existing?.cost?.input,
+    output: perMillion(info.output_cost_per_token) ?? existing?.cost?.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; it publishes real values as soon as TensorX fills the field in.
+    cache_write: perMillion(info.cache_creation_input_token_cost) ?? existing?.cost?.cache_write,
+  };
+
+  const toolFlags = [info.supports_tool_choice, info.supports_function_calling]
+    .filter((flag) => flag !== null && flag !== undefined);
+
+  // `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: info.supports_vision ?? existing?.attachment,
+    reasoning: info.supports_reasoning ?? existing?.reasoning,
+    tool_call: toolFlags.length > 0 ? toolFlags.some(Boolean) : existing?.tool_call,
+    cost,
+    limit,
+  };
+  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;
+}

From 22c26d26491ce9a8c65031028ecabf521886acff Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 06:47:05 +0200
Subject: [PATCH 02/14] chore(tensorx): move reasoning wire paths into leading
 TOML headers

The sync re-serializes provider files through Bun.TOML.parse, which
discards comments; only the leading block is re-attached. Trailing
comments on `type = "toggle"` / `type = "effort"` lines were therefore
lost on the first sync run.

AGENTS.md requires a leading top-of-file comment carrying the exact wire
path for every toggle, so the request syntax moves there and survives
future runs. No data changes.
---
 providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml   | 4 ++--
 providers/tensorx/models/deepseek/deepseek-r1-0528.toml     | 3 ++-
 providers/tensorx/models/deepseek/deepseek-v3.2.toml        | 4 ++--
 .../tensorx/models/deepseek/deepseek-v4-flash-0731.toml     | 5 +++--
 providers/tensorx/models/deepseek/deepseek-v4-flash.toml    | 3 ++-
 providers/tensorx/models/deepseek/deepseek-v4-pro.toml      | 3 ++-
 providers/tensorx/models/minimax/minimax-m2.5.toml          | 3 ++-
 providers/tensorx/models/minimax/minimax-m3.toml            | 3 ++-
 providers/tensorx/models/moonshotai/kimi-k2.6.toml          | 3 ++-
 providers/tensorx/models/moonshotai/kimi-k2.7-code.toml     | 3 ++-
 providers/tensorx/models/moonshotai/kimi-k3.toml            | 5 +++--
 .../tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml   | 3 ++-
 providers/tensorx/models/openai/gpt-oss-120b.toml           | 3 ++-
 providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml        | 3 ++-
 providers/tensorx/models/qwen/qwen3.5-9b.toml               | 4 ++--
 providers/tensorx/models/z-ai/glm-4.7.toml                  | 3 ++-
 providers/tensorx/models/z-ai/glm-5-turbo.toml              | 6 ++++--
 providers/tensorx/models/z-ai/glm-5.1.toml                  | 6 ++++--
 providers/tensorx/models/z-ai/glm-5.2.toml                  | 6 ++++--
 providers/tensorx/models/z-ai/glm-5.toml                    | 3 ++-
 providers/tensorx/models/z-ai/glm-5v-turbo.toml             | 6 ++++--
 21 files changed, 52 insertions(+), 30 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
index 93143b88c6..34f6124bc1 100644
--- a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
+++ b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
@@ -1,3 +1,4 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 name = "DeepSeek Chat V3.1"
 description = "DeepSeek chat model for instruction following, coding, and analysis"
 family = "deepseek"
@@ -10,9 +11,8 @@ tool_call = true
 knowledge = "2024-11"
 open_weights = true
 
-
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
index 5584722dd5..9a30b0a01f 100644
--- a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
+++ b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
@@ -1,10 +1,11 @@
+# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
 base_model = "deepseek/deepseek-r1"
 name = "DeepSeek R1-0528"
 release_date = "2025-05-28"
 last_updated = "2025-05-28"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; reasoning is mandatory, "none" is rejected
+type = "effort"
 values = ["minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/deepseek/deepseek-v3.2.toml b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
index dc3a096e9e..348e51b711 100644
--- a/providers/tensorx/models/deepseek/deepseek-v3.2.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
@@ -1,3 +1,4 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 name = "DeepSeek V3.2"
 description = "DeepSeek chat model for instruction following, coding, and analysis"
 family = "deepseek"
@@ -10,9 +11,8 @@ tool_call = true
 knowledge = "2025-05"
 open_weights = true
 
-
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
index 5c7bc69ad3..e775f90948 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
 base_model = "deepseek/deepseek-v4-flash-0731"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
 
 [cost]
 input = 0.25
@@ -9,4 +10,4 @@ output = 0.30
 cache_read = 0.06
 
 [limit]
-context = 1048576
\ No newline at end of file
+context = 1048576
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
index fdea8b9a03..aa2068bd2b 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
 base_model = "deepseek/deepseek-v4-flash"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
 
 [cost]
 input = 0.15
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
index ebe06937ea..71d6f837ba 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default off)
 base_model = "deepseek/deepseek-v4-pro"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": true}} (default off)
+type = "toggle"
 
 [cost]
 input = 1.75
diff --git a/providers/tensorx/models/minimax/minimax-m2.5.toml b/providers/tensorx/models/minimax/minimax-m2.5.toml
index b9c42b9646..af9e83dc13 100644
--- a/providers/tensorx/models/minimax/minimax-m2.5.toml
+++ b/providers/tensorx/models/minimax/minimax-m2.5.toml
@@ -1,7 +1,8 @@
+# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
 base_model = "minimax/MiniMax-M2.5"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; reasoning is mandatory, "none" is rejected
+type = "effort"
 values = ["minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/minimax/minimax-m3.toml b/providers/tensorx/models/minimax/minimax-m3.toml
index d9b2cb7dcb..fd320e2283 100644
--- a/providers/tensorx/models/minimax/minimax-m3.toml
+++ b/providers/tensorx/models/minimax/minimax-m3.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking_mode = enabled|disabled (default "adaptive")
 base_model = "minimax/MiniMax-M3"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking_mode": "enabled" | "disabled"}} (default "adaptive")
+type = "toggle"
 
 [cost]
 input = 0.4
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.6.toml b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
index 4490b5bfe1..128f94094d 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.6.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default on)
 base_model = "moonshotai/kimi-k2.6"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": false}} (default on)
+type = "toggle"
 
 [cost]
 input = 1
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
index 77034868c0..1a404e46d6 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
@@ -1,7 +1,8 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default on)
 base_model = "moonshotai/kimi-k2.7-code"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": false}} (default on)
+type = "toggle"
 
 [cost]
 input = 1.25
diff --git a/providers/tensorx/models/moonshotai/kimi-k3.toml b/providers/tensorx/models/moonshotai/kimi-k3.toml
index c5f9a35187..403c0b14a0 100644
--- a/providers/tensorx/models/moonshotai/kimi-k3.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k3.toml
@@ -1,10 +1,11 @@
 # AI SDK auto-discovers native IDs via /v1/models — field values map those IDs
+# Toggle: chat_template_kwargs.thinking = true|false (default on)
 base_model = "moonshotai/kimi-k3"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"thinking": false}} (default on)
+type = "toggle"
 
 [cost]
 input = 3.00
 output = 15.00
-cache_read = 0.75
\ No newline at end of file
+cache_read = 0.75
diff --git a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
index 279eb86ea6..8423a4f38a 100644
--- a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
+++ b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
@@ -1,7 +1,8 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/openai/gpt-oss-120b.toml b/providers/tensorx/models/openai/gpt-oss-120b.toml
index b34a24534e..b6a30e839c 100644
--- a/providers/tensorx/models/openai/gpt-oss-120b.toml
+++ b/providers/tensorx/models/openai/gpt-oss-120b.toml
@@ -1,8 +1,9 @@
+# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
 base_model = "openai/gpt-oss-120b"
 knowledge = "2024-10"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; reasoning is mandatory, "none" is rejected
+type = "effort"
 values = ["minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
index 063495e1fb..fadb6545fa 100644
--- a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
@@ -1,7 +1,8 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "alibaba/qwen3.5-122b-a10b"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/qwen/qwen3.5-9b.toml b/providers/tensorx/models/qwen/qwen3.5-9b.toml
index 93d4f8b02f..6419ca38b1 100644
--- a/providers/tensorx/models/qwen/qwen3.5-9b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-9b.toml
@@ -1,8 +1,8 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "alibaba/qwen3.5-9b"
 
-
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-4.7.toml b/providers/tensorx/models/z-ai/glm-4.7.toml
index 931c963e6a..c33966c544 100644
--- a/providers/tensorx/models/z-ai/glm-4.7.toml
+++ b/providers/tensorx/models/z-ai/glm-4.7.toml
@@ -1,7 +1,8 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-4.7"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-5-turbo.toml b/providers/tensorx/models/z-ai/glm-5-turbo.toml
index 4e557e1c0c..9cde43e494 100644
--- a/providers/tensorx/models/z-ai/glm-5-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5-turbo.toml
@@ -1,10 +1,12 @@
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-5-turbo"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": false}} (default on)
+type = "toggle"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-5.1.toml b/providers/tensorx/models/z-ai/glm-5.1.toml
index 869157ad21..1ff0297c3a 100644
--- a/providers/tensorx/models/z-ai/glm-5.1.toml
+++ b/providers/tensorx/models/z-ai/glm-5.1.toml
@@ -1,10 +1,12 @@
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-5.1"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": false}} (default on)
+type = "toggle"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-5.2.toml b/providers/tensorx/models/z-ai/glm-5.2.toml
index 04350a241e..2aef9e70ff 100644
--- a/providers/tensorx/models/z-ai/glm-5.2.toml
+++ b/providers/tensorx/models/z-ai/glm-5.2.toml
@@ -1,10 +1,12 @@
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-5.2"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": false}} (default on)
+type = "toggle"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index 6d547a7361..5feee3affa 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -1,7 +1,8 @@
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-5"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]
diff --git a/providers/tensorx/models/z-ai/glm-5v-turbo.toml b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
index c16e5d0679..ff12f7d1a7 100644
--- a/providers/tensorx/models/z-ai/glm-5v-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
@@ -1,10 +1,12 @@
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# Effort: reasoning_effort = <value>; "none" disables reasoning
 base_model = "zhipuai/glm-5v-turbo"
 
 [[reasoning_options]]
-type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": false}} (default on)
+type = "toggle"
 
 [[reasoning_options]]
-type = "effort" # API: {"reasoning_effort": <value>}; "none" disables reasoning
+type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 [cost]

From 7f2df40304d3a704060d32f5354b144d82610da2 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 06:47:26 +0200
Subject: [PATCH 03/14] chore(tensorx): sync output limits and cache pricing
 from own API

Source: https://api.tensorx.ai/v1/model/info

- max_output_tokens is published per model and was previously missing or
  stale on ten entries (deepseek-v4-*, kimi-k2.6/k2.7-code/k3, glm-5.1,
  glm-5.2, minimax-m3, qwen3.5-9b)
- deepseek-v4-flash-0731 cache_read corrected 0.06 -> 0.0625

cache_creation_input_token_cost is null for every model in the catalog,
so cache_write is left at its authored value rather than being dropped:
a null there means "not published", not "not charged".
---
 .../tensorx/models/deepseek/deepseek-v4-flash-0731.toml    | 7 ++++---
 providers/tensorx/models/deepseek/deepseek-v4-flash.toml   | 3 ++-
 providers/tensorx/models/deepseek/deepseek-v4-pro.toml     | 3 ++-
 providers/tensorx/models/minimax/minimax-m3.toml           | 4 ++--
 providers/tensorx/models/moonshotai/kimi-k2.6.toml         | 3 +++
 providers/tensorx/models/moonshotai/kimi-k2.7-code.toml    | 3 +++
 providers/tensorx/models/moonshotai/kimi-k3.toml           | 7 +++++--
 providers/tensorx/models/qwen/qwen3.5-9b.toml              | 3 +++
 providers/tensorx/models/z-ai/glm-5.1.toml                 | 4 ++--
 providers/tensorx/models/z-ai/glm-5.2.toml                 | 3 ++-
 10 files changed, 28 insertions(+), 12 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
index e775f90948..8c7ae2bdb7 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
@@ -6,8 +6,9 @@ type = "toggle"
 
 [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
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
index aa2068bd2b..934e1d58fe 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
@@ -11,4 +11,5 @@ cache_read = 0.0375
 cache_write = 0.1875
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
index 71d6f837ba..48919a414a 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
@@ -11,4 +11,5 @@ cache_read = 0.4375
 cache_write = 2.185
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000
diff --git a/providers/tensorx/models/minimax/minimax-m3.toml b/providers/tensorx/models/minimax/minimax-m3.toml
index fd320e2283..1bb0971570 100644
--- a/providers/tensorx/models/minimax/minimax-m3.toml
+++ b/providers/tensorx/models/minimax/minimax-m3.toml
@@ -10,5 +10,5 @@ output = 2
 cache_read = 0.1
 
 [limit]
-context = 1048576
-output = 131072
+context = 1_048_576
+output = 64_000
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.6.toml b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
index 128f94094d..b27b2030b1 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.6.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
@@ -9,3 +9,6 @@ input = 1
 output = 4
 cache_read = 0.25
 cache_write = 1.25
+
+[limit]
+output = 64_000
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
index 1a404e46d6..a19e4ac666 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
@@ -8,3 +8,6 @@ type = "toggle"
 input = 1.25
 output = 4.5
 cache_read = 0.3125
+
+[limit]
+output = 64_000
diff --git a/providers/tensorx/models/moonshotai/kimi-k3.toml b/providers/tensorx/models/moonshotai/kimi-k3.toml
index 403c0b14a0..0ed892c0fe 100644
--- a/providers/tensorx/models/moonshotai/kimi-k3.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k3.toml
@@ -6,6 +6,9 @@ base_model = "moonshotai/kimi-k3"
 type = "toggle"
 
 [cost]
-input = 3.00
-output = 15.00
+input = 3
+output = 15
 cache_read = 0.75
+
+[limit]
+output = 64_000
diff --git a/providers/tensorx/models/qwen/qwen3.5-9b.toml b/providers/tensorx/models/qwen/qwen3.5-9b.toml
index 6419ca38b1..c3d0bd7e62 100644
--- a/providers/tensorx/models/qwen/qwen3.5-9b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-9b.toml
@@ -10,3 +10,6 @@ input = 0.15
 output = 0.2
 cache_read = 0.0375
 cache_write = 0.1875
+
+[limit]
+output = 262_144
diff --git a/providers/tensorx/models/z-ai/glm-5.1.toml b/providers/tensorx/models/z-ai/glm-5.1.toml
index 1ff0297c3a..02c82a4bbf 100644
--- a/providers/tensorx/models/z-ai/glm-5.1.toml
+++ b/providers/tensorx/models/z-ai/glm-5.1.toml
@@ -16,5 +16,5 @@ cache_read = 0.35
 cache_write = 1.75
 
 [limit]
-context = 202752
-output = 202752
+context = 202_752
+output = 64_000
diff --git a/providers/tensorx/models/z-ai/glm-5.2.toml b/providers/tensorx/models/z-ai/glm-5.2.toml
index 2aef9e70ff..a74689fe5b 100644
--- a/providers/tensorx/models/z-ai/glm-5.2.toml
+++ b/providers/tensorx/models/z-ai/glm-5.2.toml
@@ -15,4 +15,5 @@ output = 4.5
 cache_read = 0.375
 
 [limit]
-context = 1048576
+context = 1_048_576
+output = 64_000

From a732e279f8072d43f4f0eeef0a909064b8e5edb6 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 07:00:17 +0200
Subject: [PATCH 04/14] fix(tensorx): declare reasoning side channel and
 kimi-k2.5 thinking toggle

---
 providers/tensorx/models/deepseek/deepseek-r1-0528.toml   | 3 +++
 providers/tensorx/models/deepseek/deepseek-v3.2.toml      | 3 +++
 .../tensorx/models/deepseek/deepseek-v4-flash-0731.toml   | 3 +++
 providers/tensorx/models/deepseek/deepseek-v4-flash.toml  | 3 +++
 providers/tensorx/models/deepseek/deepseek-v4-pro.toml    | 3 +++
 providers/tensorx/models/minimax/minimax-m2.5.toml        | 3 +++
 providers/tensorx/models/minimax/minimax-m3.toml          | 3 +++
 providers/tensorx/models/moonshotai/kimi-k2.5.toml        | 8 +++++++-
 providers/tensorx/models/moonshotai/kimi-k2.6.toml        | 3 +++
 providers/tensorx/models/moonshotai/kimi-k2.7-code.toml   | 3 +++
 providers/tensorx/models/moonshotai/kimi-k3.toml          | 3 +++
 providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml      | 3 +++
 providers/tensorx/models/qwen/qwen3.5-9b.toml             | 3 +++
 providers/tensorx/models/z-ai/glm-5-turbo.toml            | 3 +++
 providers/tensorx/models/z-ai/glm-5.1.toml                | 3 +++
 providers/tensorx/models/z-ai/glm-5.2.toml                | 3 +++
 providers/tensorx/models/z-ai/glm-5.toml                  | 3 +++
 providers/tensorx/models/z-ai/glm-5v-turbo.toml           | 3 +++
 18 files changed, 58 insertions(+), 1 deletion(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
index 9a30b0a01f..9d2e14ef42 100644
--- a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
+++ b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
@@ -8,6 +8,9 @@ last_updated = "2025-05-28"
 type = "effort"
 values = ["minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.66
 output = 2.6
diff --git a/providers/tensorx/models/deepseek/deepseek-v3.2.toml b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
index 348e51b711..491bf277d9 100644
--- a/providers/tensorx/models/deepseek/deepseek-v3.2.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
@@ -15,6 +15,9 @@ open_weights = true
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.3
 output = 0.5
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
index 8c7ae2bdb7..cea139857e 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
@@ -4,6 +4,9 @@ base_model = "deepseek/deepseek-v4-flash-0731"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.25
 output = 0.3
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
index 934e1d58fe..e0e19b5041 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
@@ -4,6 +4,9 @@ base_model = "deepseek/deepseek-v4-flash"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.15
 output = 0.3
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
index 48919a414a..ae9d34f09a 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
@@ -4,6 +4,9 @@ base_model = "deepseek/deepseek-v4-pro"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.75
 output = 3.5
diff --git a/providers/tensorx/models/minimax/minimax-m2.5.toml b/providers/tensorx/models/minimax/minimax-m2.5.toml
index af9e83dc13..ff39dd12db 100644
--- a/providers/tensorx/models/minimax/minimax-m2.5.toml
+++ b/providers/tensorx/models/minimax/minimax-m2.5.toml
@@ -5,6 +5,9 @@ base_model = "minimax/MiniMax-M2.5"
 type = "effort"
 values = ["minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.3
 output = 1.2
diff --git a/providers/tensorx/models/minimax/minimax-m3.toml b/providers/tensorx/models/minimax/minimax-m3.toml
index 1bb0971570..48eff16652 100644
--- a/providers/tensorx/models/minimax/minimax-m3.toml
+++ b/providers/tensorx/models/minimax/minimax-m3.toml
@@ -4,6 +4,9 @@ base_model = "minimax/MiniMax-M3"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.4
 output = 2
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.5.toml b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
index 83821146bc..e14ff53bbe 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.5.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
@@ -1,5 +1,11 @@
+# Toggle: chat_template_kwargs.thinking = true|false (default on)
 base_model = "moonshotai/kimi-k2.5"
-reasoning_options = []
+
+[[reasoning_options]]
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
 
 [cost]
 input = 0.5
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.6.toml b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
index b27b2030b1..4eb9dbd777 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.6.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
@@ -4,6 +4,9 @@ base_model = "moonshotai/kimi-k2.6"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1
 output = 4
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
index a19e4ac666..cb2a21846c 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
@@ -4,6 +4,9 @@ base_model = "moonshotai/kimi-k2.7-code"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.25
 output = 4.5
diff --git a/providers/tensorx/models/moonshotai/kimi-k3.toml b/providers/tensorx/models/moonshotai/kimi-k3.toml
index 0ed892c0fe..43cee375f6 100644
--- a/providers/tensorx/models/moonshotai/kimi-k3.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k3.toml
@@ -5,6 +5,9 @@ base_model = "moonshotai/kimi-k3"
 [[reasoning_options]]
 type = "toggle"
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 3
 output = 15
diff --git a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
index fadb6545fa..f773843a0e 100644
--- a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
@@ -5,6 +5,9 @@ base_model = "alibaba/qwen3.5-122b-a10b"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.5
 output = 3.5
diff --git a/providers/tensorx/models/qwen/qwen3.5-9b.toml b/providers/tensorx/models/qwen/qwen3.5-9b.toml
index c3d0bd7e62..79ef102b6f 100644
--- a/providers/tensorx/models/qwen/qwen3.5-9b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-9b.toml
@@ -5,6 +5,9 @@ base_model = "alibaba/qwen3.5-9b"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.15
 output = 0.2
diff --git a/providers/tensorx/models/z-ai/glm-5-turbo.toml b/providers/tensorx/models/z-ai/glm-5-turbo.toml
index 9cde43e494..b549cb0c58 100644
--- a/providers/tensorx/models/z-ai/glm-5-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5-turbo.toml
@@ -9,6 +9,9 @@ type = "toggle"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.2
 output = 4
diff --git a/providers/tensorx/models/z-ai/glm-5.1.toml b/providers/tensorx/models/z-ai/glm-5.1.toml
index 02c82a4bbf..ec2324116f 100644
--- a/providers/tensorx/models/z-ai/glm-5.1.toml
+++ b/providers/tensorx/models/z-ai/glm-5.1.toml
@@ -9,6 +9,9 @@ type = "toggle"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.4
 output = 4.4
diff --git a/providers/tensorx/models/z-ai/glm-5.2.toml b/providers/tensorx/models/z-ai/glm-5.2.toml
index a74689fe5b..dae334ea32 100644
--- a/providers/tensorx/models/z-ai/glm-5.2.toml
+++ b/providers/tensorx/models/z-ai/glm-5.2.toml
@@ -9,6 +9,9 @@ type = "toggle"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.5
 output = 4.5
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index 5feee3affa..550415c357 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -5,6 +5,9 @@ base_model = "zhipuai/glm-5"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1
 output = 3.2
diff --git a/providers/tensorx/models/z-ai/glm-5v-turbo.toml b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
index ff12f7d1a7..e945cb3d2e 100644
--- a/providers/tensorx/models/z-ai/glm-5v-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
@@ -9,6 +9,9 @@ type = "toggle"
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1.2
 output = 4

From a81d094e138b7653be013ef408c2378308040e49 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 07:42:55 +0200
Subject: [PATCH 05/14] fix(tensorx): honor local base_model, correct GLM
 effort controls

---
 packages/core/src/sync/providers/tensorx.ts         | 13 ++++++++++++-
 .../models/deepseek/deepseek-v4-flash-0731.toml     |  3 +++
 .../tensorx/models/deepseek/deepseek-v4-flash.toml  |  3 +++
 .../tensorx/models/deepseek/deepseek-v4-pro.toml    |  3 +++
 providers/tensorx/models/z-ai/glm-5-turbo.toml      | 10 ++++------
 providers/tensorx/models/z-ai/glm-5.1.toml          | 10 ++++------
 providers/tensorx/models/z-ai/glm-5.2.toml          |  8 +++-----
 providers/tensorx/models/z-ai/glm-5v-turbo.toml     | 10 ++++------
 8 files changed, 36 insertions(+), 24 deletions(-)

diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
index ab0e02f61c..86f5c26d39 100644
--- a/packages/core/src/sync/providers/tensorx.ts
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -60,6 +60,13 @@ export const tensorx = {
       `Add the lab entry under \`models/\` (or hand-author the provider file) first. 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) {
@@ -86,7 +93,11 @@ export const tensorx = {
     if (model.model_info.mode !== "chat") return undefined;
 
     const existing = context.existing(model.model_name);
-    const baseModel = resolveBaseModel(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.
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
index cea139857e..043fc04bf4 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
@@ -1,4 +1,7 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted here but not honored: low/high/max and an
+# invalid value all produce the same reasoning-token distribution, so this
+# host is toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-flash-0731"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
index e0e19b5041..71a4d392c8 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
@@ -1,4 +1,7 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted here but not honored: low/high/max and an
+# invalid value all produce the same reasoning-token distribution, so this
+# host is toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-flash"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
index ae9d34f09a..b3bff278c6 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
@@ -1,4 +1,7 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
+# reasoning_effort is accepted here but not honored: low/high/max and an
+# invalid value all produce the same reasoning-token distribution, so this
+# host is toggle-only despite the lab exposing effort high|max.
 base_model = "deepseek/deepseek-v4-pro"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/z-ai/glm-5-turbo.toml b/providers/tensorx/models/z-ai/glm-5-turbo.toml
index b549cb0c58..686e8f6b20 100644
--- a/providers/tensorx/models/z-ai/glm-5-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5-turbo.toml
@@ -1,13 +1,11 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
+# `none` already expresses it, so effort is the single declared control.
+# Effort: reasoning_effort = none|low|medium|high (verified; others rejected 400)
 base_model = "zhipuai/glm-5-turbo"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "low", "medium", "high"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.1.toml b/providers/tensorx/models/z-ai/glm-5.1.toml
index ec2324116f..281a6b268d 100644
--- a/providers/tensorx/models/z-ai/glm-5.1.toml
+++ b/providers/tensorx/models/z-ai/glm-5.1.toml
@@ -1,13 +1,11 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
+# `none` already expresses it, so effort is the single declared control.
+# Effort: reasoning_effort = none|low|medium|high|max (verified; others rejected 400)
 base_model = "zhipuai/glm-5.1"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "low", "medium", "high", "max"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.2.toml b/providers/tensorx/models/z-ai/glm-5.2.toml
index dae334ea32..8601224b0b 100644
--- a/providers/tensorx/models/z-ai/glm-5.2.toml
+++ b/providers/tensorx/models/z-ai/glm-5.2.toml
@@ -1,10 +1,8 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
+# `none` already expresses it, so effort is the single declared control.
+# Effort: reasoning_effort = none|minimal|low|medium|high|xhigh|max (verified; others rejected 400)
 base_model = "zhipuai/glm-5.2"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
 values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
diff --git a/providers/tensorx/models/z-ai/glm-5v-turbo.toml b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
index e945cb3d2e..ac2ba3cd3e 100644
--- a/providers/tensorx/models/z-ai/glm-5v-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
@@ -1,13 +1,11 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
+# `none` already expresses it, so effort is the single declared control.
+# Effort: reasoning_effort = none|low|medium|high (verified; others rejected 400)
 base_model = "zhipuai/glm-5v-turbo"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "low", "medium", "high"]
 
 [interleaved]
 field = "reasoning_content"

From 56807afd6b4342b7355c76bbdbf3ec2627c45a39 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 08:37:51 +0200
Subject: [PATCH 06/14] fix(tensorx): guard partial pricing and author measured
 reasoning controls

---
 packages/core/src/sync/providers/tensorx.ts   | 38 +++++++++++--------
 .../tensorx/models/moonshotai/kimi-k2.5.toml  |  8 ++--
 .../tensorx/models/moonshotai/kimi-k2.6.toml  |  2 +
 .../models/moonshotai/kimi-k2.7-code.toml     |  2 +
 .../tensorx/models/moonshotai/kimi-k3.toml    |  6 +++
 .../tensorx/models/z-ai/glm-5-turbo.toml      |  9 ++---
 providers/tensorx/models/z-ai/glm-5.1.toml    |  9 ++---
 providers/tensorx/models/z-ai/glm-5.2.toml    | 11 ++++--
 providers/tensorx/models/z-ai/glm-5.toml      |  6 ++-
 .../tensorx/models/z-ai/glm-5v-turbo.toml     |  9 ++---
 10 files changed, 60 insertions(+), 40 deletions(-)

diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
index 86f5c26d39..c7fa301aa7 100644
--- a/packages/core/src/sync/providers/tensorx.ts
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -103,10 +103,8 @@ export const tensorx = {
     // hand-authoring instead of being invented.
     if (baseModel === undefined && existing === undefined) return undefined;
 
-    return {
-      id: model.model_name,
-      model: buildTensorXModel(model, baseModel, existing),
-    };
+    const built = buildTensorXModel(model, baseModel, existing);
+    return built === undefined ? undefined : { id: model.model_name, model: built };
   },
 } satisfies SyncProvider<TensorXModel>;
 
@@ -114,7 +112,7 @@ function buildTensorXModel(
   model: TensorXModel,
   baseModel: string | undefined,
   existing: ExistingModel | undefined,
-): SyncedModel {
+): SyncedModel | undefined {
   const info = model.model_info;
 
   const limit = {
@@ -122,16 +120,26 @@ function buildTensorXModel(
     output: info.max_output_tokens ?? existing?.limit?.output,
   };
 
-  const cost = {
-    ...existing?.cost,
-    input: perMillion(info.input_cost_per_token) ?? existing?.cost?.input,
-    output: perMillion(info.output_cost_per_token) ?? existing?.cost?.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; it publishes real values as soon as TensorX fills the field in.
-    cache_write: perMillion(info.cache_creation_input_token_cost) ?? existing?.cost?.cache_write,
-  };
+  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);
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.5.toml b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
index e14ff53bbe..871007b6f9 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.5.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
@@ -1,8 +1,8 @@
-# Toggle: chat_template_kwargs.thinking = true|false (default on)
+# Neither control is honored here: chat_template_kwargs.thinking and
+# reasoning_effort (including none, and an invalid value) all leave the
+# reasoning-token count unchanged over 8 runs. Reasoning is always on.
 base_model = "moonshotai/kimi-k2.5"
-
-[[reasoning_options]]
-type = "toggle"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.6.toml b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
index 4eb9dbd777..1457ff3269 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.6.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
@@ -1,4 +1,6 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default on)
+# reasoning_effort is accepted but not honored here — an invalid value and
+# none both leave the reasoning-token count unchanged — so this is toggle-only.
 base_model = "moonshotai/kimi-k2.6"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
index cb2a21846c..2456682013 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
@@ -1,4 +1,6 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default on)
+# reasoning_effort is accepted but not honored here — an invalid value and
+# none both leave the reasoning-token count unchanged — so this is toggle-only.
 base_model = "moonshotai/kimi-k2.7-code"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/moonshotai/kimi-k3.toml b/providers/tensorx/models/moonshotai/kimi-k3.toml
index 43cee375f6..8b903afe89 100644
--- a/providers/tensorx/models/moonshotai/kimi-k3.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k3.toml
@@ -1,10 +1,16 @@
 # AI SDK auto-discovers native IDs via /v1/models — field values map those IDs
 # Toggle: chat_template_kwargs.thinking = true|false (default on)
+# Effort: reasoning_effort = low|high|max; invalid values return 400 and low
+# measurably lowers the reasoning-token count. Matches first-party Moonshot.
 base_model = "moonshotai/kimi-k3"
 
 [[reasoning_options]]
 type = "toggle"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
 [interleaved]
 field = "reasoning_content"
 
diff --git a/providers/tensorx/models/z-ai/glm-5-turbo.toml b/providers/tensorx/models/z-ai/glm-5-turbo.toml
index 686e8f6b20..7237603bb0 100644
--- a/providers/tensorx/models/z-ai/glm-5-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5-turbo.toml
@@ -1,11 +1,10 @@
-# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
-# `none` already expresses it, so effort is the single declared control.
-# Effort: reasoning_effort = none|low|medium|high (verified; others rejected 400)
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# No reasoning tokens were observed on this host under any setting, so the
+# toggle is carried from first-party z.ai rather than verified here.
 base_model = "zhipuai/glm-5-turbo"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "low", "medium", "high"]
+type = "toggle"
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.1.toml b/providers/tensorx/models/z-ai/glm-5.1.toml
index 281a6b268d..5378fc7c25 100644
--- a/providers/tensorx/models/z-ai/glm-5.1.toml
+++ b/providers/tensorx/models/z-ai/glm-5.1.toml
@@ -1,11 +1,10 @@
-# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
-# `none` already expresses it, so effort is the single declared control.
-# Effort: reasoning_effort = none|low|medium|high|max (verified; others rejected 400)
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# reasoning_effort is accepted but its graded levels were not separable here;
+# first-party z.ai exposes only the toggle, so that is the declared control.
 base_model = "zhipuai/glm-5.1"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "low", "medium", "high", "max"]
+type = "toggle"
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.2.toml b/providers/tensorx/models/z-ai/glm-5.2.toml
index 8601224b0b..5bc8c2f958 100644
--- a/providers/tensorx/models/z-ai/glm-5.2.toml
+++ b/providers/tensorx/models/z-ai/glm-5.2.toml
@@ -1,11 +1,14 @@
-# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
-# `none` already expresses it, so effort is the single declared control.
-# Effort: reasoning_effort = none|minimal|low|medium|high|xhigh|max (verified; others rejected 400)
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# Effort: reasoning_effort = high|max — z.ai maps none|minimal to off,
+# low|medium to high and xhigh to max, so high|max are the effective levels.
 base_model = "zhipuai/glm-5.2"
 
+[[reasoning_options]]
+type = "toggle"
+
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["high", "max"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index 550415c357..fc34500580 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -1,9 +1,11 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Effort: reasoning_effort = none|high|max; none verified to disable reasoning.
+# chat_template_kwargs.enable_thinking is accepted but not honored here, and
+# the graded levels were not separable, so only z.ai's effective levels list.
 base_model = "zhipuai/glm-5"
 
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "high", "max"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5v-turbo.toml b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
index ac2ba3cd3e..cc93bd2dda 100644
--- a/providers/tensorx/models/z-ai/glm-5v-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
@@ -1,11 +1,10 @@
-# Off is effort=none; chat_template_kwargs.enable_thinking also toggles, but
-# `none` already expresses it, so effort is the single declared control.
-# Effort: reasoning_effort = none|low|medium|high (verified; others rejected 400)
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
+# No reasoning tokens were observed on this host under any setting, so the
+# toggle is carried from first-party z.ai rather than verified here.
 base_model = "zhipuai/glm-5v-turbo"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "low", "medium", "high"]
+type = "toggle"
 
 [interleaved]
 field = "reasoning_content"

From d49ed5c86d71ffb6b5956fd3db79da89f017197f Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 09:52:24 +0200
Subject: [PATCH 07/14] fix(tensorx): publish only measured GLM-5 on/off, stop
 auto-creating reasoners

---
 packages/core/src/sync/providers/tensorx.ts | 10 ++++++++--
 providers/tensorx/models/z-ai/glm-5.toml    | 11 ++++++-----
 2 files changed, 14 insertions(+), 7 deletions(-)

diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
index c7fa301aa7..ce7d827e8d 100644
--- a/packages/core/src/sync/providers/tensorx.ts
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -42,6 +42,12 @@ 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
@@ -56,8 +62,8 @@ export const tensorx = {
   skippedNotice(ids) {
     if (ids.length === 0) return [];
     return [
-      `${ids.length} TensorX chat models were not created because /v1/model/info publishes no name, description, or release date to author a model from.`,
-      `Add the lab entry under \`models/\` (or hand-author the provider file) first. Skipped remote IDs: ${ids.map((id) => `\`${id}\``).join(", ")}`,
+      `${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) {
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index fc34500580..dea547aa61 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -1,11 +1,12 @@
-# Effort: reasoning_effort = none|high|max; none verified to disable reasoning.
-# chat_template_kwargs.enable_thinking is accepted but not honored here, and
-# the graded levels were not separable, so only z.ai's effective levels list.
+# Toggle: reasoning_effort = none turns reasoning off (verified: 0 reasoning
+# tokens); any other accepted value leaves it on. chat_template_kwargs.
+# enable_thinking is accepted but is a no-op here, and no graded level was
+# separable from another, so only on/off is published. First-party z.ai is
+# likewise toggle-only for GLM-5.
 base_model = "zhipuai/glm-5"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "high", "max"]
+type = "toggle"
 
 [interleaved]
 field = "reasoning_content"

From b318da2cc3409cef7a47cb9754fafe85aa0486ef Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 10:14:02 +0200
Subject: [PATCH 08/14] fix(tensorx): verify GLM turbo reasoning controls via
 reasoning_content

---
 providers/tensorx/models/z-ai/glm-5-turbo.toml  | 9 ++++++---
 providers/tensorx/models/z-ai/glm-5v-turbo.toml | 9 ++++++---
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/providers/tensorx/models/z-ai/glm-5-turbo.toml b/providers/tensorx/models/z-ai/glm-5-turbo.toml
index 7237603bb0..9f444b466d 100644
--- a/providers/tensorx/models/z-ai/glm-5-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5-turbo.toml
@@ -1,6 +1,9 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# No reasoning tokens were observed on this host under any setting, so the
-# toggle is carried from first-party z.ai rather than verified here.
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on).
+# Verified: false empties reasoning_content, true and the default fill it.
+# This host streams reasoning into reasoning_content but does not count it in
+# usage.reasoning_tokens, which reads 0 even while reasoning runs — measure
+# the field, not the token count. reasoning_effort = none also suppresses it,
+# but no graded level was separable, so only on/off is published.
 base_model = "zhipuai/glm-5-turbo"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/z-ai/glm-5v-turbo.toml b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
index cc93bd2dda..b884c3f19d 100644
--- a/providers/tensorx/models/z-ai/glm-5v-turbo.toml
+++ b/providers/tensorx/models/z-ai/glm-5v-turbo.toml
@@ -1,6 +1,9 @@
-# Toggle: chat_template_kwargs.enable_thinking = true|false (default on)
-# No reasoning tokens were observed on this host under any setting, so the
-# toggle is carried from first-party z.ai rather than verified here.
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default on).
+# Verified: false empties reasoning_content, true and the default fill it.
+# This host streams reasoning into reasoning_content but does not count it in
+# usage.reasoning_tokens, which reads 0 even while reasoning runs — measure
+# the field, not the token count. reasoning_effort = none also suppresses it,
+# but no graded level was separable, so only on/off is published.
 base_model = "zhipuai/glm-5v-turbo"
 
 [[reasoning_options]]

From ae50adabc169972b62a09fed74397908c3169bc0 Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 10:49:55 +0200
Subject: [PATCH 09/14] fix(tensorx): re-verify reasoning controls against
 reasoning_content

---
 .../models/deepseek/deepseek-v4-flash-0731.toml     |  7 ++++---
 .../tensorx/models/deepseek/deepseek-v4-flash.toml  |  7 ++++---
 .../tensorx/models/deepseek/deepseek-v4-pro.toml    |  7 ++++---
 providers/tensorx/models/moonshotai/kimi-k2.5.toml  |  7 ++++---
 providers/tensorx/models/moonshotai/kimi-k2.6.toml  |  7 ++++---
 .../tensorx/models/moonshotai/kimi-k2.7-code.toml   |  7 ++++---
 providers/tensorx/models/moonshotai/kimi-k3.toml    | 13 ++++++-------
 providers/tensorx/models/z-ai/glm-5.toml            |  9 +++++----
 8 files changed, 35 insertions(+), 29 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
index 043fc04bf4..c109b56717 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash-0731.toml
@@ -1,7 +1,8 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
-# reasoning_effort is accepted here but not honored: low/high/max and an
-# invalid value all produce the same reasoning-token distribution, so this
-# host is toggle-only despite the lab exposing effort high|max.
+# 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]]
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
index 71a4d392c8..08fa020600 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-flash.toml
@@ -1,7 +1,8 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
-# reasoning_effort is accepted here but not honored: low/high/max and an
-# invalid value all produce the same reasoning-token distribution, so this
-# host is toggle-only despite the lab exposing effort high|max.
+# 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]]
diff --git a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
index b3bff278c6..fec75b2d13 100644
--- a/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v4-pro.toml
@@ -1,7 +1,8 @@
 # Toggle: chat_template_kwargs.thinking = true|false (default off)
-# reasoning_effort is accepted here but not honored: low/high/max and an
-# invalid value all produce the same reasoning-token distribution, so this
-# host is toggle-only despite the lab exposing effort high|max.
+# 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]]
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.5.toml b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
index 871007b6f9..db3656d067 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.5.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.5.toml
@@ -1,6 +1,7 @@
-# Neither control is honored here: chat_template_kwargs.thinking and
-# reasoning_effort (including none, and an invalid value) all leave the
-# reasoning-token count unchanged over 8 runs. Reasoning is always on.
+# No caller control. Measured by reasoning_content length over 6 runs each,
+# chat_template_kwargs.thinking = false and reasoning_effort = none both leave
+# reasoning_content filled (0/6 empty). usage.reasoning_tokens is not the
+# signal here because this host under-reports it. Reasoning is always on.
 base_model = "moonshotai/kimi-k2.5"
 reasoning_options = []
 
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.6.toml b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
index 1457ff3269..3d926729dc 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.6.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.6.toml
@@ -1,6 +1,7 @@
-# Toggle: chat_template_kwargs.thinking = true|false (default on)
-# reasoning_effort is accepted but not honored here — an invalid value and
-# none both leave the reasoning-token count unchanged — so this is toggle-only.
+# Toggle: chat_template_kwargs.thinking = true|false (default on); verified by
+# reasoning_content going empty when false.
+# reasoning_effort is accepted but not honored: none, low, high, max and an
+# invalid value all leave reasoning_content in the same range.
 base_model = "moonshotai/kimi-k2.6"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
index 2456682013..e9088bbfec 100644
--- a/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k2.7-code.toml
@@ -1,6 +1,7 @@
-# Toggle: chat_template_kwargs.thinking = true|false (default on)
-# reasoning_effort is accepted but not honored here — an invalid value and
-# none both leave the reasoning-token count unchanged — so this is toggle-only.
+# Toggle: chat_template_kwargs.thinking = true|false (default on); verified by
+# reasoning_content going empty when false.
+# reasoning_effort is accepted but not honored: none, low, high, max and an
+# invalid value all leave reasoning_content in the same range.
 base_model = "moonshotai/kimi-k2.7-code"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/moonshotai/kimi-k3.toml b/providers/tensorx/models/moonshotai/kimi-k3.toml
index 8b903afe89..4985c54953 100644
--- a/providers/tensorx/models/moonshotai/kimi-k3.toml
+++ b/providers/tensorx/models/moonshotai/kimi-k3.toml
@@ -1,15 +1,14 @@
 # AI SDK auto-discovers native IDs via /v1/models — field values map those IDs
-# Toggle: chat_template_kwargs.thinking = true|false (default on)
-# Effort: reasoning_effort = low|high|max; invalid values return 400 and low
-# measurably lowers the reasoning-token count. Matches first-party Moonshot.
+# Effort: reasoning_effort = none|low|high|max, measured by reasoning_content
+# length. none empties it in 9/9 runs and depth grows low < high < max;
+# invalid values return 400. chat_template_kwargs.thinking = false is
+# unreliable here (reasoning still streamed in 3/9 runs), so none is the
+# published off switch and no toggle is declared.
 base_model = "moonshotai/kimi-k3"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["low", "high", "max"]
+values = ["none", "low", "high", "max"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index dea547aa61..03db1e3f3b 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -1,7 +1,8 @@
-# Toggle: reasoning_effort = none turns reasoning off (verified: 0 reasoning
-# tokens); any other accepted value leaves it on. chat_template_kwargs.
-# enable_thinking is accepted but is a no-op here, and no graded level was
-# separable from another, so only on/off is published. First-party z.ai is
+# Toggle: reasoning_effort = none turns reasoning off — verified by
+# reasoning_content going empty, not by usage.reasoning_tokens, which this
+# host under-reports. Any other accepted value leaves it on.
+# chat_template_kwargs.enable_thinking is accepted but is a no-op here, and no
+# graded level was separable, so only on/off is published. First-party z.ai is
 # likewise toggle-only for GLM-5.
 base_model = "zhipuai/glm-5"
 

From aefa5824e20f91b213da53159ff9cf60731b69ae Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 11:24:46 +0200
Subject: [PATCH 10/14] fix(tensorx): model effort-as-off as effort, keep
 attachment and modalities coherent

---
 packages/core/src/sync/providers/tensorx.ts   | 19 ++++++++++++++++++-
 .../models/deepseek/deepseek-v3.2.toml        |  8 ++++++--
 .../models/qwen/qwen3.5-122b-a10b.toml        |  8 ++++++--
 providers/tensorx/models/z-ai/glm-5.toml      | 14 +++++++-------
 4 files changed, 37 insertions(+), 12 deletions(-)

diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
index ce7d827e8d..bff889152f 100644
--- a/packages/core/src/sync/providers/tensorx.ts
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -150,15 +150,32 @@ function buildTensorXModel(
   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. Authoring one without the
+  // other would publish `attachment = true` beside a text-only modality list, or
+  // `attachment = false` while image input stays inherited. Only `image` is
+  // touched — the flag says nothing about video or pdf.
+  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"),
+    };
+
   // `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: info.supports_vision ?? existing?.attachment,
+    attachment: vision ?? 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;
diff --git a/providers/tensorx/models/deepseek/deepseek-v3.2.toml b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
index 491bf277d9..874da8eb6c 100644
--- a/providers/tensorx/models/deepseek/deepseek-v3.2.toml
+++ b/providers/tensorx/models/deepseek/deepseek-v3.2.toml
@@ -1,4 +1,8 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# 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"
@@ -13,7 +17,7 @@ open_weights = true
 
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "high"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
index f773843a0e..60fb5c6822 100644
--- a/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-122b-a10b.toml
@@ -1,9 +1,13 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Effort: reasoning_effort = none|high. Off is the effort value none, not a
+# boolean field: chat_template_kwargs.enable_thinking and thinking are both
+# accepted but are no-ops here. Measured by reasoning_content length (this
+# host under-reports usage.reasoning_tokens); none empties it, invalid values
+# return 400, and no graded level was separable, so one on level is published.
 base_model = "alibaba/qwen3.5-122b-a10b"
 
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "high"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-5.toml b/providers/tensorx/models/z-ai/glm-5.toml
index 03db1e3f3b..0696922a62 100644
--- a/providers/tensorx/models/z-ai/glm-5.toml
+++ b/providers/tensorx/models/z-ai/glm-5.toml
@@ -1,13 +1,13 @@
-# Toggle: reasoning_effort = none turns reasoning off — verified by
-# reasoning_content going empty, not by usage.reasoning_tokens, which this
-# host under-reports. Any other accepted value leaves it on.
-# chat_template_kwargs.enable_thinking is accepted but is a no-op here, and no
-# graded level was separable, so only on/off is published. First-party z.ai is
-# likewise toggle-only for GLM-5.
+# Effort: reasoning_effort = none|high. Off is the effort value none, not a
+# boolean field: chat_template_kwargs.enable_thinking is accepted but is a no-op
+# on this host, so no toggle is declared. Verified by reasoning_content going
+# empty for none, not by usage.reasoning_tokens, which this host under-reports.
+# No graded level was separable from another, so one on level is published.
 base_model = "zhipuai/glm-5"
 
 [[reasoning_options]]
-type = "toggle"
+type = "effort"
+values = ["none", "high"]
 
 [interleaved]
 field = "reasoning_content"

From 417aaf528f4ae3322fc2e7e99e5a62e02aaaac8e Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 11:32:44 +0200
Subject: [PATCH 11/14] fix(tensorx): narrow remaining effort enums to measured
 or lab-baseline controls

---
 .../tensorx/models/deepseek/deepseek-chat-v3.1.toml    |  6 +++++-
 providers/tensorx/models/minimax/minimax-m2.5.toml     | 10 +++++-----
 .../models/nvidia/nemotron-3-super-120b-a12b.toml      |  6 +++++-
 providers/tensorx/models/openai/gpt-oss-120b.toml      |  8 ++++++--
 providers/tensorx/models/qwen/qwen3.5-9b.toml          |  8 ++++++--
 providers/tensorx/models/z-ai/glm-4.7.toml             | 10 +++++++---
 6 files changed, 34 insertions(+), 14 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
index 34f6124bc1..1270bd036e 100644
--- a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
+++ b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
@@ -1,4 +1,8 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Not reachable with the sync key (403, models=['public']), so nothing below
+# could be measured against this host. Controls are carried from the lab/peer
+# baseline; [interleaved] is deliberately omitted rather than assumed, even
+# though all 18 reachable models here stream reasoning_content.
+# No lab or peer entry authors reasoning options, so the list is left as-is.
 name = "DeepSeek Chat V3.1"
 description = "DeepSeek chat model for instruction following, coding, and analysis"
 family = "deepseek"
diff --git a/providers/tensorx/models/minimax/minimax-m2.5.toml b/providers/tensorx/models/minimax/minimax-m2.5.toml
index ff39dd12db..3fa9dbfe7f 100644
--- a/providers/tensorx/models/minimax/minimax-m2.5.toml
+++ b/providers/tensorx/models/minimax/minimax-m2.5.toml
@@ -1,9 +1,9 @@
-# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
+# No caller control. reasoning_effort = none is rejected (400), low/medium/
+# high/max overlap entirely over 4 runs each, and chat_template_kwargs
+# enable_thinking and thinking are both no-ops. Measured by reasoning_content
+# length. Reasoning is always on, matching the lab and OpenRouter entries.
 base_model = "minimax/MiniMax-M2.5"
-
-[[reasoning_options]]
-type = "effort"
-values = ["minimal", "low", "medium", "high", "xhigh", "max"]
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
index 8423a4f38a..c248acbc3a 100644
--- a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
+++ b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
@@ -1,4 +1,8 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Not reachable with the sync key (403, models=['public']), so nothing below
+# could be measured against this host. Controls are carried from the lab/peer
+# baseline; [interleaved] is deliberately omitted rather than assumed, even
+# though all 18 reachable models here stream reasoning_content.
+# No lab or peer entry authors reasoning options, so the list is left as-is.
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
 [[reasoning_options]]
diff --git a/providers/tensorx/models/openai/gpt-oss-120b.toml b/providers/tensorx/models/openai/gpt-oss-120b.toml
index b6a30e839c..dfcf2045ea 100644
--- a/providers/tensorx/models/openai/gpt-oss-120b.toml
+++ b/providers/tensorx/models/openai/gpt-oss-120b.toml
@@ -1,10 +1,14 @@
-# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
+# Not reachable with the sync key (403, models=['public']), so nothing below
+# could be measured against this host. Controls are carried from the lab/peer
+# baseline; [interleaved] is deliberately omitted rather than assumed, even
+# though all 18 reachable models here stream reasoning_content.
+# OpenRouter, the closest same-surface peer, exposes low|medium|high.
 base_model = "openai/gpt-oss-120b"
 knowledge = "2024-10"
 
 [[reasoning_options]]
 type = "effort"
-values = ["minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["low", "medium", "high"]
 
 [cost]
 input = 0.04
diff --git a/providers/tensorx/models/qwen/qwen3.5-9b.toml b/providers/tensorx/models/qwen/qwen3.5-9b.toml
index 79ef102b6f..0f6d4f879f 100644
--- a/providers/tensorx/models/qwen/qwen3.5-9b.toml
+++ b/providers/tensorx/models/qwen/qwen3.5-9b.toml
@@ -1,9 +1,13 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Effort: reasoning_effort = none|high. Off is the effort value none, not a
+# boolean field: chat_template_kwargs.enable_thinking and thinking are no-ops
+# here. Measured by reasoning_content length; none empties it, invalid values
+# return 400, and over 4 runs each low/medium/high/max overlap entirely
+# (means within 60 chars), so one on level is published.
 base_model = "alibaba/qwen3.5-9b"
 
 [[reasoning_options]]
 type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+values = ["none", "high"]
 
 [interleaved]
 field = "reasoning_content"
diff --git a/providers/tensorx/models/z-ai/glm-4.7.toml b/providers/tensorx/models/z-ai/glm-4.7.toml
index c33966c544..e0fcce5b2a 100644
--- a/providers/tensorx/models/z-ai/glm-4.7.toml
+++ b/providers/tensorx/models/z-ai/glm-4.7.toml
@@ -1,9 +1,13 @@
-# Effort: reasoning_effort = <value>; "none" disables reasoning
+# Not reachable with the sync key (403, models=['public']), so nothing below
+# could be measured against this host. Controls are carried from the lab/peer
+# baseline; [interleaved] is deliberately omitted rather than assumed, even
+# though all 18 reachable models here stream reasoning_content.
+# First-party z.ai GLM-4.7 is toggle-only.
+# Toggle: chat_template_kwargs.enable_thinking = true|false
 base_model = "zhipuai/glm-4.7"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+type = "toggle"
 
 [cost]
 input = 0.6

From eca4d608c29053f386f3f434ecf3858141b8cd4d Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 11:39:37 +0200
Subject: [PATCH 12/14] fix(tensorx): copy peer toggle for unreachable DeepSeek
 V3.1 and Nemotron

---
 .../models/deepseek/deepseek-chat-v3.1.toml       | 15 ++++++++-------
 .../models/nvidia/nemotron-3-super-120b-a12b.toml | 15 ++++++++-------
 2 files changed, 16 insertions(+), 14 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
index 1270bd036e..f75eace0bf 100644
--- a/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
+++ b/providers/tensorx/models/deepseek/deepseek-chat-v3.1.toml
@@ -1,8 +1,10 @@
-# Not reachable with the sync key (403, models=['public']), so nothing below
-# could be measured against this host. Controls are carried from the lab/peer
-# baseline; [interleaved] is deliberately omitted rather than assumed, even
-# though all 18 reachable models here stream reasoning_content.
-# No lab or peer entry authors reasoning options, so the list is left as-is.
+# 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"
@@ -16,8 +18,7 @@ knowledge = "2024-11"
 open_weights = true
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+type = "toggle"
 
 [cost]
 input = 0.2
diff --git a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
index c248acbc3a..b310722164 100644
--- a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
+++ b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
@@ -1,13 +1,14 @@
-# Not reachable with the sync key (403, models=['public']), so nothing below
-# could be measured against this host. Controls are carried from the lab/peer
-# baseline; [interleaved] is deliberately omitted rather than assumed, even
-# though all 18 reachable models here stream reasoning_content.
-# No lab or peer entry authors reasoning options, so the list is left as-is.
+# Not reachable with the sync key (403, models=['public']), so nothing here
+# could be measured against this host. First-party providers/nvidia/models/
+# nvidia/nemotron-3-super-120b-a12b.toml authors a toggle, so that control
+# shape is copied rather than keeping a full effort enum.
+# Toggle: wire path is neither documented by the lab entry nor verifiable with
+# this key — confirm it before relying on it.
+# [interleaved] is omitted rather than assumed, for the same reason.
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
 [[reasoning_options]]
-type = "effort"
-values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
+type = "toggle"
 
 [cost]
 input = 0.3

From a960ce2db464c27c96f5aea35feb97c2a20c031a Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 13:58:17 +0200
Subject: [PATCH 13/14] fix(tensorx): match peer baseline for DeepSeek R1-0528
 reasoning options

---
 providers/tensorx/models/deepseek/deepseek-r1-0528.toml | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
index 9d2e14ef42..2d2925531f 100644
--- a/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
+++ b/providers/tensorx/models/deepseek/deepseek-r1-0528.toml
@@ -1,12 +1,11 @@
-# Effort: reasoning_effort = <value>; reasoning is mandatory, "none" is rejected
+# 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]]
-type = "effort"
-values = ["minimal", "low", "medium", "high", "xhigh", "max"]
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"

From e2e920f8f99fb2a9850d9340301d07135ffc879b Mon Sep 17 00:00:00 2001
From: bu6n <b@u6n.ch>
Date: Tue, 4 Aug 2026 14:57:50 +0200
Subject: [PATCH 14/14] fix(tensorx): document Nemotron toggle wire path,
 derive attachment from modalities

---
 packages/core/src/sync/providers/tensorx.ts        | 14 +++++++++-----
 .../models/nvidia/nemotron-3-super-120b-a12b.toml  | 14 +++++++-------
 2 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/packages/core/src/sync/providers/tensorx.ts b/packages/core/src/sync/providers/tensorx.ts
index bff889152f..a8b71dcd0e 100644
--- a/packages/core/src/sync/providers/tensorx.ts
+++ b/packages/core/src/sync/providers/tensorx.ts
@@ -151,10 +151,10 @@ function buildTensorXModel(
     .filter((flag) => flag !== null && flag !== undefined);
 
   // supports_vision is the only modality signal the catalog carries, so
-  // `attachment` and `modalities.input` move together. Authoring one without the
-  // other would publish `attachment = true` beside a text-only modality list, or
-  // `attachment = false` while image input stays inherited. Only `image` is
-  // touched — the flag says nothing about video or pdf.
+  // `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
@@ -165,12 +165,16 @@ function buildTensorXModel(
         ? (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: vision ?? existing?.attachment,
+    attachment,
     reasoning: info.supports_reasoning ?? existing?.reasoning,
     tool_call: toolFlags.length > 0 ? toolFlags.some(Boolean) : existing?.tool_call,
     cost,
diff --git a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
index b310722164..770b6699ed 100644
--- a/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
+++ b/providers/tensorx/models/nvidia/nemotron-3-super-120b-a12b.toml
@@ -1,10 +1,10 @@
-# Not reachable with the sync key (403, models=['public']), so nothing here
-# could be measured against this host. First-party providers/nvidia/models/
-# nvidia/nemotron-3-super-120b-a12b.toml authors a toggle, so that control
-# shape is copied rather than keeping a full effort enum.
-# Toggle: wire path is neither documented by the lab entry nor verifiable with
-# this key — confirm it before relying on it.
-# [interleaved] is omitted rather than assumed, for the same reason.
+# Toggle: chat_template_kwargs.enable_thinking = true|false
+# Documented for this model by Workers AI, another OpenAI-compatible host:
+# https://developers.cloudflare.com/workers-ai/models/nemotron-3-120b-a12b/sync-input.json
+# It is also the field every GLM model on this host uses. Not reachable with the
+# sync key (403, models=['public']), so it could not be confirmed here; the
+# control shape follows first-party providers/nvidia/models/nvidia/
+# nemotron-3-super-120b-a12b.toml. [interleaved] is omitted rather than assumed.
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
 [[reasoning_options]]