pi-commandcode-provider/tests/models/models.test.ts

import assert from "node:assert/strict"
import { test } from "node:test"

import {
  accountApiBase,
  apiForModelId,
  baseUrlForApi,
  fetchLiveCatalog,
  getModelsTimeoutMs,
  modelsFromCatalog,
  modelsFromLive,
  parseLiveCatalog,
  providerHeaders,
  thinkingLevelMapFor,
  toProviderModel,
} from "../../src/models.ts"

const PROVIDER_API_BASE = "https://api.commandcode.ai/provider/v1"

/** Shape of the real GET /provider/v1/models response. */
const liveCatalogResponse = {
  object: "list",
  data: [
    { id: "deepseek/deepseek-v4.1-flash", object: "model", name: "DeepSeek V4.1 Flash", context_length: 1_000_000 },
    { id: "claude-sonnet-4-6", object: "model", name: "Claude Sonnet 4.6", context_length: 1_000_000 },
    { id: "vendor/brand-new-model", object: "model", name: "Brand New", context_length: 32_768 },
  ],
}

test("parseLiveCatalog reads id, name and context window", () => {
  const models = parseLiveCatalog(liveCatalogResponse)

  assert.deepEqual(
    models.map((model) => model.id),
    ["deepseek/deepseek-v4.1-flash", "claude-sonnet-4-6", "vendor/brand-new-model"],
  )
  assert.equal(models[2]?.contextWindow, 32_768)
})

test("parseLiveCatalog rejects malformed catalogs", () => {
  assert.throws(() => parseLiveCatalog({ object: "list", data: [] }), /empty model catalog/)
  assert.throws(() => parseLiveCatalog({ object: "collection", data: [{}] }), /'list'/)
  assert.throws(
    () => parseLiveCatalog({ object: "list", data: [{ id: "x", name: "X" }] }),
    /context_length/,
  )
})

test("modelsFromLive uses CLI metadata when the model is known", () => {
  const [deepseek, claude, unknown] = modelsFromLive(parseLiveCatalog(liveCatalogResponse))

  assert.equal(deepseek?.reasoning, true)
  assert.deepEqual(deepseek?.efforts, ["low", "high", "max"])
  assert.deepEqual(deepseek?.input, ["text", "image"])
  assert.equal(deepseek?.cost.output, 0.6)
  assert.equal(claude?.api, "anthropic-messages")

  // Unknown models stay usable but text-only and unpriced until the catalog syncs.
  assert.equal(unknown?.reasoning, false)
  assert.deepEqual(unknown?.input, ["text"])
  assert.deepEqual(unknown?.cost, { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 })
  assert.equal(unknown?.maxTokens, 32_768)
})

test("modelsFromCatalog exposes every generated entry", async () => {
  const { CATALOG } = await import("../../src/catalog.ts")
  const models = modelsFromCatalog()

  assert.equal(models.length, CATALOG.length)
  assert.ok(models.every((model) => model.maxTokens > 0 && model.contextWindow > 0))
})

test("baseline models without a published context window fall back to a usable default", () => {
  // The CLI reference lists GLM-5.1 as "—" for context, and pi cannot run a model with a 0 window.
  const glm = modelsFromCatalog().find((model) => model.id === "zai-org/GLM-5.1")

  assert.equal(glm?.contextWindow, 200_000)
  assert.ok((glm?.maxTokens ?? 0) > 0)
})

test("modelsFromLive prefers the live context window over the CLI snapshot", () => {
  const [model] = modelsFromLive([
    { id: "deepseek/deepseek-v4.1-flash", name: "DeepSeek V4.1 Flash", contextWindow: 512_000 },
  ])

  assert.equal(model?.contextWindow, 512_000)
  assert.equal(model?.maxTokens, 65_536)
  assert.deepEqual(model?.input, ["text", "image"])
})

test("live context windows and DeepSeek V4.1 vision/effort metadata survive the merge", () => {
  const [model] = modelsFromLive([
    { id: "deepseek/deepseek-v4.1-flash", name: "DeepSeek V4.1 Flash", contextWindow: 1_000_000 },
  ])
  assert.ok(model)
  const config = toProviderModel(model, PROVIDER_API_BASE)

  // Vision and the opt-in max effort are what the API actually serves for V4.1.
  assert.deepEqual(config.input, ["text", "image"])
  assert.deepEqual(config.thinkingLevelMap, {
    minimal: null,
    low: "low",
    medium: null,
    high: "high",
    xhigh: null,
    max: "max",
  })
  assert.equal(
    (config.compat as { supportsReasoningEffort?: boolean }).supportsReasoningEffort,
    true,
  )
})

test("api and base URL follow the model family", () => {
  assert.equal(apiForModelId("claude-opus-5"), "anthropic-messages")
  assert.equal(apiForModelId("deepseek/deepseek-v4.1-flash"), "openai-completions")
  assert.equal(baseUrlForApi(PROVIDER_API_BASE, "openai-completions"), PROVIDER_API_BASE)
  // pi appends /v1/messages to the Anthropic base URL.
  assert.equal(baseUrlForApi(PROVIDER_API_BASE, "anthropic-messages"), "https://api.commandcode.ai/provider")
})

test("accountApiBase strips the provider namespace", () => {
  assert.equal(accountApiBase(PROVIDER_API_BASE), "https://api.commandcode.ai")
  assert.equal(accountApiBase("https://example.test/provider/v1/"), "https://example.test")
})

test("thinkingLevelMap hides levels the model does not offer", () => {
  assert.deepEqual(thinkingLevelMapFor(["low", "high", "max"]), {
    minimal: null,
    low: "low",
    medium: null,
    high: "high",
    xhigh: null,
    max: "max",
  })
})

test("toProviderModel maps a Claude model onto the Anthropic adapter", () => {
  const [claude] = modelsFromLive(parseLiveCatalog(liveCatalogResponse)).slice(1)
  assert.ok(claude)
  const config = toProviderModel(claude, PROVIDER_API_BASE)

  assert.equal(config.api, "anthropic-messages")
  assert.equal(config.baseUrl, "https://api.commandcode.ai/provider")
  assert.equal(config.reasoning, true)
  assert.deepEqual(config.thinkingLevelMap?.high, "high")
  assert.equal(
    (config.compat as { forceAdaptiveThinking?: boolean }).forceAdaptiveThinking,
    true,
  )
  assert.equal(config.cost.cacheWrite, 3.75)
})

test("toProviderModel maps an OpenAI-compatible model onto the completions adapter", () => {
  const [deepseek] = modelsFromLive(parseLiveCatalog(liveCatalogResponse))
  assert.ok(deepseek)
  const config = toProviderModel(deepseek, PROVIDER_API_BASE)

  assert.equal(config.api, "openai-completions")
  assert.equal(config.baseUrl, PROVIDER_API_BASE)
  assert.deepEqual(config.compat, {
    supportsStore: false,
    supportsDeveloperRole: false,
    supportsReasoningEffort: true,
    maxTokensField: "max_tokens",
  })
  assert.deepEqual(config.input, ["text", "image"])
})

test("fetchLiveCatalog parses the response and surfaces HTTP failures", async () => {
  const ok = await fetchLiveCatalog({
    fetchImpl: async () => new Response(JSON.stringify(liveCatalogResponse), { status: 200 }),
  })
  assert.equal(ok.length, 3)

  await assert.rejects(
    fetchLiveCatalog({ fetchImpl: async () => new Response("nope", { status: 503 }) }),
    /503/,
  )
})

test("environment overrides for base URL, headers and timeout", () => {
  assert.deepEqual(providerHeaders({ CMD_ZDR: "1" }), { "x-cmd-zdr": "1" })
  assert.equal(providerHeaders({ COMMANDCODE_ZDR: "1" })?.["x-cmd-zdr"], "1")
  assert.equal(providerHeaders({}), undefined)
  assert.equal(getModelsTimeoutMs({ COMMANDCODE_MODELS_TIMEOUT_MS: "2500" }), 2500)
  assert.equal(getModelsTimeoutMs({ COMMANDCODE_MODELS_TIMEOUT_MS: "-1" }), 10_000)
})