feat(venice): add reasoning options

packages/core/script/generate-venice.ts

@@ -9,6 +9,42 @@ import { ModelFamilyValues } from "../src/family.js";
 const API_ENDPOINT = "https://api.venice.ai/api/v1/models?type=text";
 
 // Zod schemas for API response validation
+const ReasoningEffort = z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]);
+type ReasoningOption = {
+  type: "effort";
+  values: Array<z.infer<typeof ReasoningEffort>>;
+};
+
+const effort = (...values: ReasoningOption["values"]): ReasoningOption[] => [{ type: "effort", values }];
+
+// Venice documents these model-specific values even where /models is stale or incomplete.
+// Source: https://docs.venice.ai/guides/features/reasoning-models
+export const REASONING_OVERRIDES: Record<string, ReasoningOption[]> = {
+  "claude-opus-4-5": effort("low", "medium", "high"),
+  "claude-opus-4-6": effort("low", "medium", "high", "max"),
+  "claude-opus-4-6-fast": effort("low", "medium", "high", "max"),
+  "claude-sonnet-4-5": effort("low", "medium", "high"),
+  "claude-sonnet-4-6": effort("low", "medium", "high"),
+  "gemini-3-flash-preview": effort("minimal", "low", "medium", "high"),
+  "kimi-k2-5": effort("low", "medium", "high"),
+  "openai-gpt-52": effort("none", "low", "medium", "high", "xhigh"),
+  "openai-gpt-52-codex": effort("low", "medium", "high", "xhigh"),
+  "openai-gpt-53-codex": effort("low", "medium", "high", "xhigh"),
+  "qwen3-5-35b-a3b": effort("low", "medium", "high"),
+  "zai-org-glm-5-1": [],
+
+  // Provisional until funded Venice probes can confirm that its proxy preserves
+  // OpenAI's current controls. Sources:
+  // https://developers.openai.com/api/docs/models/gpt-5.4
+  // https://developers.openai.com/api/docs/models/gpt-5.4-pro
+  // https://developers.openai.com/api/docs/guides/reasoning
+  "openai-gpt-54": effort("none", "low", "medium", "high", "xhigh"),
+  "openai-gpt-54-mini": effort("none", "low", "medium", "high", "xhigh"),
+  "openai-gpt-54-pro": effort("medium", "high", "xhigh"),
+  "openai-gpt-55": effort("none", "low", "medium", "high", "xhigh"),
+  "openai-gpt-55-pro": effort("medium", "high", "xhigh"),
+};
+
 const Capabilities = z
   .object({
     optimizedForCode: z.boolean().optional(),
@@ -17,6 +53,9 @@ const Capabilities = z
     supportsFunctionCalling: z.boolean().optional(),
     supportsLogProbs: z.boolean().optional(),
     supportsReasoning: z.boolean().optional(),
+    supportsReasoningEffort: z.boolean().optional(),
+    reasoningEffortOptions: z.array(ReasoningEffort).optional(),
+    defaultReasoningEffort: ReasoningEffort.optional(),
     supportsResponseSchema: z.boolean().optional(),
     supportsVideoInput: z.boolean().optional(),
     supportsVision: z.boolean().optional(),
@@ -142,6 +181,7 @@ interface ExistingModel {
   family?: string;
   attachment?: boolean;
   reasoning?: boolean;
+  reasoning_options?: ReasoningOption[];
   tool_call?: boolean;
   structured_output?: boolean;
   temperature?: boolean;
@@ -235,6 +275,7 @@ interface MergedModel {
   family?: string;
   attachment: boolean;
   reasoning: boolean;
+  reasoning_options?: ReasoningOption[];
   tool_call: boolean;
   structured_output?: boolean;
   temperature: boolean;
@@ -267,9 +308,10 @@ interface MergedModel {
   };
 }
 
-function mergeModel(
+export function mergeModel(
   apiModel: z.infer<typeof VeniceModel>,
   existing: ExistingModel | null,
+  reportDiscrepancy: (message: string) => void = console.warn,
 ): MergedModel {
   const spec = apiModel.model_spec;
   const caps = spec.capabilities;
@@ -309,6 +351,31 @@ function mergeModel(
     },
   };
 
+  const override = REASONING_OVERRIDES[apiModel.id];
+  const catalogOptions = caps.supportsReasoningEffort === true && caps.reasoningEffortOptions !== undefined
+    ? effort(...caps.reasoningEffortOptions)
+    : undefined;
+  const curatedOptions = existing?.reasoning_options;
+  const selectedOptions = override ?? curatedOptions ?? (existing === null ? catalogOptions : undefined);
+
+  if (selectedOptions !== undefined) {
+    merged.reasoning_options = selectedOptions;
+  }
+
+  const catalogClaim = catalogOptions ?? (caps.supportsReasoningEffort === false ? [] : undefined);
+  if (override !== undefined && JSON.stringify(override) !== JSON.stringify(curatedOptions) && curatedOptions !== undefined) {
+    reportDiscrepancy(`${apiModel.id}: documented override replaces curated reasoning_options`);
+  }
+  if (
+    selectedOptions !== undefined &&
+    catalogClaim !== undefined &&
+    JSON.stringify(selectedOptions) !== JSON.stringify(catalogClaim)
+  ) {
+    reportDiscrepancy(
+      `${apiModel.id}: preserving ${override !== undefined ? "documented override" : "curated reasoning_options"} despite catalog ${caps.supportsReasoningEffort === false ? "supportsReasoningEffort=false" : "option mismatch"}`,
+    );
+  }
+
   // structured_output only if true
   if (caps.supportsResponseSchema === true) {
     merged.structured_output = true;
@@ -352,7 +419,7 @@ function mergeModel(
   return merged;
 }
 
-function formatToml(model: MergedModel): string {
+export function formatToml(model: MergedModel): string {
   const lines: string[] = [];
 
   // Basic fields
@@ -362,6 +429,9 @@ function formatToml(model: MergedModel): string {
   }
   lines.push(`attachment = ${model.attachment}`);
   lines.push(`reasoning = ${model.reasoning}`);
+  if (model.reasoning_options?.length === 0) {
+    lines.push("reasoning_options = []");
+  }
   lines.push(`tool_call = ${model.tool_call}`);
   if (model.structured_output !== undefined) {
     lines.push(`structured_output = ${model.structured_output}`);
@@ -377,6 +447,13 @@ function formatToml(model: MergedModel): string {
     lines.push(`status = "${model.status}"`);
   }
 
+  for (const option of model.reasoning_options ?? []) {
+    lines.push("");
+    lines.push("[[reasoning_options]]");
+    lines.push(`type = "${option.type}"`);
+    lines.push(`values = [${option.values.map((value) => `"${value}"`).join(", ")}]`);
+  }
+
   // Interleaved section (if present)
   if (model.interleaved !== undefined) {
     lines.push("");
@@ -437,7 +514,7 @@ interface Changes {
   newValue: string;
 }
 
-function detectChanges(
+export function detectChanges(
   existing: ExistingModel | null,
   merged: MergedModel,
 ): Changes[] {
@@ -459,7 +536,7 @@ function detectChanges(
 
   const formatValue = (val: unknown): string => {
     if (typeof val === "number") return formatNumber(val);
-    if (Array.isArray(val)) return `[${val.join(", ")}]`;
+    if (Array.isArray(val)) return JSON.stringify(val);
     if (val === undefined) return "(none)";
     return String(val);
   };
@@ -468,6 +545,7 @@ function detectChanges(
   compare("family", existing.family, merged.family);
   compare("attachment", existing.attachment, merged.attachment);
   compare("reasoning", existing.reasoning, merged.reasoning);
+  compare("reasoning_options", existing.reasoning_options, merged.reasoning_options);
   compare("tool_call", existing.tool_call, merged.tool_call);
   compare("structured_output", existing.structured_output, merged.structured_output);
   compare("open_weights", existing.open_weights, merged.open_weights);
@@ -650,4 +728,6 @@ async function main() {
   }
 }
 
-await main();
+if (import.meta.main) {
+  await main();
+}

packages/core/test/venice-generator.test.ts

@@ -0,0 +1,121 @@
+import { expect, test } from "bun:test";
+
+import {
+  detectChanges,
+  formatToml,
+  mergeModel,
+  REASONING_OVERRIDES,
+} from "../script/generate-venice.js";
+
+type Effort = "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
+const options = (...values: Effort[]) => [{ type: "effort" as const, values }];
+
+function model(id: string, capabilities: Record<string, unknown>) {
+  return {
+    created: 1_700_000_000,
+    id,
+    model_spec: {
+      availableContextTokens: 128_000,
+      maxCompletionTokens: 32_000,
+      capabilities: { supportsReasoning: true, ...capabilities },
+      name: "Test Model",
+    },
+    object: "model",
+    owned_by: "venice.ai",
+    type: "text",
+  };
+}
+
+test("curated options survive false and stale Venice catalog metadata", () => {
+  const claude = options("low", "medium", "high", "max");
+  const codex = options("low", "medium", "high", "xhigh");
+  const discrepancies: string[] = [];
+
+  expect(mergeModel(model("claude-opus-4-7", {
+    supportsReasoningEffort: false,
+  }), { reasoning_options: claude }, discrepancies.push.bind(discrepancies)).reasoning_options).toEqual(claude);
+  expect(mergeModel(model("openai-gpt-56-codex", {
+    supportsReasoningEffort: true,
+    reasoningEffortOptions: ["none", "low"],
+  }), { reasoning_options: codex }, discrepancies.push.bind(discrepancies)).reasoning_options).toEqual(codex);
+  expect(discrepancies).toHaveLength(2);
+});
+
+test("documented override beats stale curated and catalog options", () => {
+  const discrepancies: string[] = [];
+  const merged = mergeModel(model("openai-gpt-52", {
+    supportsReasoningEffort: true,
+    reasoningEffortOptions: ["minimal", "low", "high"],
+  }), { reasoning_options: options("low", "high") }, discrepancies.push.bind(discrepancies));
+
+  expect(merged.reasoning_options).toEqual(options("none", "low", "medium", "high", "xhigh"));
+  expect(discrepancies).toHaveLength(2);
+});
+
+test("catalog fills only an uncurated new model", () => {
+  const merged = mergeModel(model("new-reasoner", {
+    supportsReasoningEffort: true,
+    reasoningEffortOptions: ["low", "high"],
+  }), null);
+
+  expect(merged.reasoning_options).toEqual(options("low", "high"));
+});
+
+test("catalog does not fill an unresolved existing model", () => {
+  const existing = { reasoning: true };
+  const merged = mergeModel(model("existing-unresolved-reasoner", {
+    supportsReasoningEffort: true,
+    reasoningEffortOptions: ["low", "high"],
+  }), existing);
+
+  expect(merged.reasoning_options).toBeUndefined();
+  expect(detectChanges(existing, merged).find((change) => change.field === "reasoning_options")).toBeUndefined();
+});
+
+test("catalog false without curated evidence leaves options undefined", () => {
+  const merged = mergeModel(model("unknown-fixed-reasoner", {
+    supportsReasoningEffort: false,
+  }), null);
+
+  expect(merged.reasoning_options).toBeUndefined();
+  expect(formatToml(merged)).not.toContain("reasoning_options");
+});
+
+test("explicit curated empty options remain stable", () => {
+  const existing = { reasoning_options: [] };
+  const merged = mergeModel(model("curated-fixed-reasoner", {
+    supportsReasoningEffort: false,
+  }), existing);
+
+  expect(merged.reasoning_options).toEqual([]);
+  expect(detectChanges({ ...existing, reasoning: true }, merged).find((change) => change.field === "reasoning_options")).toBeUndefined();
+  expect(formatToml(merged)).toContain("reasoning = true\nreasoning_options = []");
+});
+
+test("formatter emits nonempty options using model TOML convention", () => {
+  const merged = mergeModel(model("new-reasoner", {
+    supportsReasoningEffort: true,
+    reasoningEffortOptions: ["none", "high"],
+  }), null);
+
+  expect(formatToml(merged)).toContain(
+    '[[reasoning_options]]\ntype = "effort"\nvalues = ["none", "high"]',
+  );
+});
+
+test("official correction fixtures remain exact", () => {
+  const expected = {
+    "claude-opus-4-6": options("low", "medium", "high", "max"),
+    "openai-gpt-52": options("none", "low", "medium", "high", "xhigh"),
+    "openai-gpt-52-codex": options("low", "medium", "high", "xhigh"),
+    "openai-gpt-54-pro": options("medium", "high", "xhigh"),
+    "gemini-3-flash-preview": options("minimal", "low", "medium", "high"),
+    "kimi-k2-5": options("low", "medium", "high"),
+    "qwen3-5-35b-a3b": options("low", "medium", "high"),
+    "zai-org-glm-5-1": [],
+  };
+
+  for (const [id, reasoningOptions] of Object.entries(expected)) {
+    expect(REASONING_OVERRIDES[id]).toEqual(reasoningOptions);
+  }
+});

providers/venice/models/aion-labs-aion-2-0.toml

@@ -2,6 +2,7 @@ name = "Aion 2.0"
 family = "o"
 attachment = false
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = false
 temperature = true
 release_date = "2026-03-24"

providers/venice/models/arcee-trinity-large-thinking.toml

@@ -2,6 +2,7 @@ name = "Trinity Large Thinking"
 family = "trinity"
 attachment = false
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/claude-opus-4-5.toml

@@ -2,6 +2,7 @@ name = "Claude Opus 4.5"
 family = "claude-opus"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/claude-opus-4-6-fast.toml

@@ -2,6 +2,7 @@ name = "Claude Opus 4.6 Fast"
 family = "claude-opus"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "max"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/claude-opus-4-6.toml

@@ -2,6 +2,7 @@ name = "Claude Opus 4.6"
 family = "claude-opus"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "max"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/claude-sonnet-4-5.toml

@@ -2,6 +2,7 @@ name = "Claude Sonnet 4.5"
 family = "claude-sonnet"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/claude-sonnet-4-6.toml

@@ -2,6 +2,7 @@ name = "Claude Sonnet 4.6"
 family = "claude-sonnet"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

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

@@ -2,6 +2,7 @@ name = "DeepSeek V3.2"
 family = "deepseek"
 attachment = false
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/gemini-3-1-pro-preview.toml

@@ -2,6 +2,7 @@ name = "Gemini 3.1 Pro Preview"
 family = "gemini-pro"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/gemini-3-5-flash.toml

@@ -2,6 +2,7 @@ name = "Gemini 3.5 Flash"
 family = "gemini-flash"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/gemini-3-flash-preview.toml

@@ -2,6 +2,7 @@ name = "Gemini 3 Flash Preview"
 family = "gemini-flash"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["minimal", "low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true

providers/venice/models/google-gemma-4-26b-a4b-it.toml

@@ -2,6 +2,7 @@ name = "Google Gemma 4 26B A4B Instruct"
 family = "gemma"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
 tool_call = true
 structured_output = true
 temperature = true