OmniRoute/tests/unit/t31-t33-t34-t38-model-specs.test.ts

import test from "node:test";
import assert from "node:assert/strict";

const { REGISTRY } = await import("../../open-sse/config/providerRegistry.ts");
const { getStaticModelsForProvider } = await import("../../src/lib/providers/staticModels.ts");
const { resolveModelAlias: resolveDeprecatedAlias } =
  await import("../../open-sse/services/modelDeprecation.ts");
const { normalizeThinkingLevel } = await import("../../open-sse/services/thinkingBudget.ts");
const {
  MODEL_SPECS,
  getModelSpec,
  capMaxOutputTokens,
  resolveModelAlias,
  getDefaultThinkingBudget,
  capThinkingBudget,
} = await import("../../src/shared/constants/modelSpecs.ts");

test("T31: antigravity static catalog exposes client-visible Gemini preview IDs", () => {
  // Antigravity exposes preview aliases to clients even though the upstream
  // still accepts its internal model identifiers.
  const staticIds = (getStaticModelsForProvider("antigravity") || []).map((m) => m.id);
  assert.ok(staticIds.includes("gemini-3-pro-preview"));
  // #3303 (agy parity, discussion #3184): the Gemini + Claude budget tiers ARE
  // client-visible on the Antigravity OAuth backend (Claude was never removed).
  assert.ok(staticIds.includes("gemini-3.1-pro-low"));
  assert.ok(staticIds.includes("claude-sonnet-4-6"));
  // The legacy cloaked Claude aliases remain absent.
  assert.ok(!staticIds.includes("gemini-claude-sonnet-4-5-thinking"));
  assert.ok(!staticIds.includes("gemini-claude-opus-4-5-thinking"));
});

test("T31: legacy Gemini aliases resolve to Gemini 3.1 IDs", () => {
  assert.equal(resolveDeprecatedAlias("gemini-3-pro-high"), "gemini-3.1-pro-high");
  assert.equal(resolveDeprecatedAlias("gemini-3-pro-low"), "gemini-3.1-pro-low");
});

test("T33: thinkingLevel string is converted into numeric thinkingBudget", () => {
  const converted = normalizeThinkingLevel({
    model: "gemini-3.1-pro-high",
    generationConfig: {
      thinkingConfig: { thinkingLevel: "HIGH" },
    },
  });

  assert.equal(converted.generationConfig.thinkingConfig.thinkingBudget, 24576);
  assert.equal(converted.generationConfig.thinkingConfig.thinkingLevel, undefined);
});

test("T34: max output tokens are capped by model spec", () => {
  assert.equal(capMaxOutputTokens("gpt-5.5", 200000), 128000);
  assert.equal(capMaxOutputTokens("gpt-5.5-xhigh", 200000), 128000);
  assert.equal(capMaxOutputTokens("gemini-3-flash", 131072), 65536);
  assert.equal(capMaxOutputTokens("gemini-3-flash"), 65536);
  assert.equal(capMaxOutputTokens("gemini-3.1-pro-high", 131072), 65535);
  assert.equal(capMaxOutputTokens("claude-opus-4-8", 200000), 128000);
  assert.equal(capMaxOutputTokens("claude-opus-4-7", 200000), 128000);
  assert.equal(capMaxOutputTokens("anthropic.claude-sonnet-4-6", 200000), 64000);
  assert.equal(capMaxOutputTokens("eu.anthropic.claude-opus-4-6", 200000), 128000);
});

test("T38: modelSpecs exposes centralized helpers with alias and prefix lookup", () => {
  assert.equal(getModelSpec("gpt-5.5").contextWindow, 1050000);
  assert.equal(getModelSpec("gpt-5.5-high").maxOutputTokens, 128000);
  assert.equal(typeof MODEL_SPECS["gemini-3.1-pro"], "object");
  assert.equal(getModelSpec("gemini-3-pro-high").maxOutputTokens, 65535);
  assert.equal(getModelSpec("gemini-3-pro-preview").maxOutputTokens, 65535);
  assert.equal(getModelSpec("gemini-3-flash-preview").maxOutputTokens, 65536);
  assert.equal(getModelSpec("gemini-3.1-pro-preview").maxOutputTokens, 65535);
  assert.equal(getModelSpec("gemini-3.1-pro-preview-customtools").maxOutputTokens, 65535);
  assert.equal(getModelSpec("claude-opus-4-7").contextWindow, 1000000);
  assert.equal(getModelSpec("claude-opus-4.8").maxOutputTokens, 128000);
  assert.equal(getModelSpec("claude-opus-4.7").maxOutputTokens, 128000);
  assert.equal(getModelSpec("bedrock/eu.anthropic.claude-sonnet-4-6").contextWindow, 1000000);
  assert.equal(getModelSpec("bedrock/anthropic.claude-sonnet-4-5").contextWindow, 200000);
  assert.equal(getModelSpec("global.anthropic.claude-opus-4-6").maxOutputTokens, 128000);
  assert.equal(resolveModelAlias("gemini-3-pro-low"), "gemini-3.1-pro-low");
  assert.equal(resolveModelAlias("gemini-3-pro-preview"), "gemini-3.1-pro");
  assert.equal(resolveModelAlias("gemini-3.1-pro-preview"), "gemini-3.1-pro");
  assert.equal(resolveModelAlias("gemini-3.1-pro-preview-customtools"), "gemini-3.1-pro");
  assert.equal(getDefaultThinkingBudget("gemini-3.1-pro-high"), 24576);
  assert.equal(capThinkingBudget("gemini-3.1-pro-low", 50000), 16000);
});

test("T38: only MiMo V2.5 and V2 Omni support vision; the *-pro/flash are text-only", () => {
  // Corrected: Xiaomi documents ONLY mimo-v2.5 and mimo-v2-omni as image-capable
  // (mimo.mi.com .../image-understanding). The *-pro chat models are text-only;
  // models.dev mislabels them (hermes-agent#18884).
  assert.equal(MODEL_SPECS["mimo-v2.5-pro"].supportsVision, false);
  assert.equal(MODEL_SPECS["mimo-v2-pro"].supportsVision, false);
  assert.equal(MODEL_SPECS["mimo-v2.5"].supportsVision, true);
  assert.equal(MODEL_SPECS["mimo-v2-omni"].supportsVision, true);
  assert.equal(MODEL_SPECS["mimo-v2-flash"].supportsVision, undefined);
});

test("opencode-go family: context/output caps match upstream provider docs", () => {
  // Qwen3.x Plus / Max: 1M context, 65K output (Bailian)
  assert.equal(getModelSpec("qwen3-max").contextWindow, 1000000);
  assert.equal(getModelSpec("qwen3-max").maxOutputTokens, 65536);
  assert.equal(getModelSpec("qwen3.7-max").contextWindow, 1000000);
  assert.equal(getModelSpec("qwen3-max-2026-01-23").contextWindow, 1000000);
  assert.equal(getModelSpec("qwen3.6-plus").contextWindow, 1000000);
  assert.equal(getModelSpec("qwen3.6-plus").maxOutputTokens, 65536);
  assert.equal(getModelSpec("qwen3.5-plus").contextWindow, 1000000);
  assert.equal(getModelSpec("qwen3.5-plus").maxOutputTokens, 65536);

  // Kimi K2.5: 262K context/output (Moonshot, parity with K2.6)
  assert.equal(getModelSpec("kimi-k2.5").contextWindow, 262144);
  assert.equal(getModelSpec("kimi-k2.5").maxOutputTokens, 262144);

  // GLM-5.x: 200K context, 128K output (Z.AI)
  assert.equal(getModelSpec("glm-5.1").contextWindow, 200000);
  assert.equal(getModelSpec("glm-5.1").maxOutputTokens, 128000);
  assert.equal(getModelSpec("glm-5").contextWindow, 200000);

  // MiniMax M2.x: ~200K context, 131K output
  assert.equal(getModelSpec("minimax-m2.7").contextWindow, 204800);
  assert.equal(getModelSpec("minimax-m2.7").maxOutputTokens, 131072);
  assert.equal(getModelSpec("minimax-m2.5").contextWindow, 200000);
  assert.equal(getModelSpec("MiniMax-M2.5").contextWindow, 200000);

  // DeepSeek V4: 1M context, 384K output
  assert.equal(getModelSpec("deepseek-v4-pro").contextWindow, 1000000);
  assert.equal(getModelSpec("deepseek-v4-pro").maxOutputTokens, 384000);
  assert.equal(getModelSpec("deepseek-v4-flash").contextWindow, 1000000);

  // Tencent Hunyuan 3 Preview: 262K context/output
  assert.equal(getModelSpec("hy3-preview").contextWindow, 262144);
  assert.equal(getModelSpec("hy3-preview").maxOutputTokens, 262144);
});

test("opencode-go family: capMaxOutputTokens grants full upstream budget", () => {
  // Without explicit specs these models would fall back to __default__ (8192).
  // Assert they now receive the real upstream cap.
  assert.equal(capMaxOutputTokens("qwen3-max", 100000), 65536);
  assert.equal(capMaxOutputTokens("qwen3.7-max", 100000), 65536);
  assert.equal(capMaxOutputTokens("qwen3-max-2026-01-23", 100000), 65536);
  assert.equal(capMaxOutputTokens("kimi-k2.5", 300000), 262144);
  assert.equal(capMaxOutputTokens("glm-5.1", 200000), 128000);
  assert.equal(capMaxOutputTokens("minimax-m2.7", 200000), 131072);
  assert.equal(capMaxOutputTokens("MiniMax-M2.5", 200000), 131072);
  assert.equal(capMaxOutputTokens("deepseek-v4-pro", 500000), 384000);
  assert.equal(capMaxOutputTokens("hy3-preview", 300000), 262144);
});