feat(core): add Modal live model backfill plugin (#47136)

packages/core/src/modal/models.ts

@@ -0,0 +1,141 @@
+export * as ModalModels from "./models.js"
+
+import { Money } from "@opencode-ai/schema/money"
+import { Option, Schema } from "effect"
+import { Model } from "../model.js"
+import { Provider } from "../provider.js"
+
+const providerID = Provider.ID.make("modal")
+
+const ReasoningOption = Schema.Struct({
+  type: Schema.Literal("effort"),
+  values: Schema.Array(Schema.NullOr(Schema.String)),
+})
+
+const RemoteModel = Schema.Struct({
+  id: Schema.String,
+  base_model_id: Schema.optional(Schema.String),
+  hugging_face_id: Schema.optional(Schema.String),
+  name: Schema.optional(Schema.String),
+  input_modalities: Schema.optional(Schema.Array(Schema.String)),
+  output_modalities: Schema.optional(Schema.Array(Schema.String)),
+  context_length: Schema.optional(Schema.Number),
+  max_output_length: Schema.optional(Schema.Number),
+  pricing: Schema.optional(
+    Schema.Struct({
+      prompt: Schema.optional(Schema.Union([Schema.String, Schema.Number])),
+      completion: Schema.optional(Schema.Union([Schema.String, Schema.Number])),
+      input_cache_read: Schema.optional(Schema.Union([Schema.String, Schema.Number])),
+    }),
+  ),
+  supported_sampling_parameters: Schema.optional(Schema.Array(Schema.String)),
+  supported_features: Schema.optional(Schema.Array(Schema.String)),
+  reasoning_options: Schema.optional(Schema.Array(ReasoningOption)),
+  interleaved: Schema.optional(
+    Schema.Union([
+      Schema.Boolean,
+      Schema.Struct({
+        field: Schema.Literals(["reasoning", "reasoning_content", "reasoning_details"]),
+      }),
+    ]),
+  ),
+})
+
+const Response = Schema.Struct({ data: Schema.Array(Schema.Unknown) })
+const decodeResponse = Schema.decodeUnknownSync(Response)
+const decodeModel = Schema.decodeUnknownOption(RemoteModel)
+
+type RemoteModel = typeof RemoteModel.Type
+
+export async function get(baseURL: string, apiKey: string, existing: readonly Model.Info[]) {
+  const response = await fetch(`${baseURL.replace(/\/+$/, "")}/models`, {
+    headers: {
+      Authorization: `Bearer ${apiKey}`,
+    },
+    signal: AbortSignal.timeout(3_000),
+  })
+  if (!response.ok) throw new Error(`Failed to fetch Modal models: ${response.status}`)
+
+  // Decode each item tolerantly so one malformed entry cannot discard the
+  // whole inventory. A malformed envelope still fails the fetch.
+  const remote = decodeResponse(await response.json()).data.flatMap((raw) => {
+    const model = Option.getOrUndefined(decodeModel(raw))
+    return model ? [model] : []
+  })
+  const templates = new Map(existing.map((model) => [model.id, model]))
+  const result = new Map<Model.ID, Model.Info>()
+  for (const item of remote) {
+    const template = templates.get(Model.ID.make(item.base_model_id ?? item.hugging_face_id ?? item.id))
+    const id = Model.ID.make(item.id)
+    result.set(id, build(id, item, baseURL, template))
+  }
+  return result
+}
+
+function price(value: string | number | undefined, fallback: Money.USDPerMillionTokens) {
+  if (value === undefined) return fallback
+  const parsed = Number(value) * 1_000_000
+  return Number.isFinite(parsed) ? Money.USDPerMillionTokens.make(parsed) : fallback
+}
+
+function limit(value: number | undefined, fallback: number) {
+  const parsed = value === undefined ? fallback : Math.trunc(value)
+  return Number.isInteger(parsed) && parsed >= 0 ? parsed : fallback
+}
+
+function build(id: Model.ID, remote: RemoteModel, baseURL: string, previous?: Model.Info) {
+  const cost = previous?.cost[0]
+  const input = previous?.limit.input
+  return Model.Info.make({
+    ...Model.Info.default(providerID, id),
+    id,
+    modelID: Model.ID.make(remote.id),
+    providerID,
+    name: remote.name ?? previous?.name ?? remote.id,
+    family: previous?.family,
+    compatibility:
+      remote.interleaved === undefined
+        ? previous?.compatibility
+        : (Model.compatibility(remote.interleaved) ?? previous?.compatibility),
+    package: Provider.aisdk("@ai-sdk/openai-compatible"),
+    settings: Provider.mergeOverlay(previous?.settings, { baseURL }),
+    headers: previous?.headers,
+    body: previous?.body,
+    capabilities: {
+      tools: remote.supported_features?.includes("tools") ?? previous?.capabilities.tools ?? true,
+      input: remote.input_modalities ?? previous?.capabilities.input ?? ["text"],
+      output: remote.output_modalities ?? previous?.capabilities.output ?? ["text"],
+    },
+    variants: remote.reasoning_options === undefined ? (previous?.variants ?? []) : variants(remote),
+    time: previous?.time ?? { released: 0 },
+    cost: [
+      {
+        input: price(remote.pricing?.prompt, cost?.input ?? Money.USDPerMillionTokens.zero),
+        output: price(remote.pricing?.completion, cost?.output ?? Money.USDPerMillionTokens.zero),
+        cache: {
+          read: price(remote.pricing?.input_cache_read, cost?.cache.read ?? Money.USDPerMillionTokens.zero),
+          write: cost?.cache.write ?? Money.USDPerMillionTokens.zero,
+        },
+      },
+    ],
+    status: previous?.status ?? "active",
+    enabled: previous?.enabled ?? true,
+    limit: {
+      context: limit(remote.context_length, previous?.limit.context ?? 0),
+      ...(input === undefined ? {} : { input }),
+      output: limit(remote.max_output_length, previous?.limit.output ?? 0),
+    },
+  })
+}
+
+function variants(remote: RemoteModel): Model.Info["variants"] {
+  const seen = new Map<string, Model.Info["variants"][number]>()
+  for (const option of remote.reasoning_options ?? []) {
+    for (const value of option.values) {
+      const effort = value ?? "none"
+      if (!seen.has(effort))
+        seen.set(effort, { id: Model.VariantID.make(effort), settings: { reasoningEffort: effort } })
+    }
+  }
+  return [...seen.values()]
+}

packages/core/src/plugin/provider.ts

@@ -15,6 +15,7 @@ import { KiloPlugin } from "./provider/kilo.js"
 import { LLMGatewayPlugin } from "./provider/llmgateway.js"
 import { LMStudioPlugin } from "./provider/lmstudio.js"
 import { MistralPlugin } from "./provider/mistral.js"
+import { ModalPlugin } from "./provider/modal.js"
 import { NvidiaPlugin } from "./provider/nvidia.js"
 import { OllamaPlugin } from "./provider/ollama.js"
 import { OpenAIPlugin } from "./provider/openai.js"
@@ -48,6 +49,7 @@ export const ProviderPlugins: PluginInternal.InternalPlugin[] = [
   LLMGatewayPlugin,
   LMStudioPlugin,
   MistralPlugin,
+  ModalPlugin,
   NvidiaPlugin,
   OllamaPlugin,
   OpencodePlugin,

packages/core/src/plugin/provider/modal.ts

@@ -0,0 +1,67 @@
+import { Effect, Semaphore, Stream } from "effect"
+import { define } from "@opencode-ai/plugin/effect/plugin"
+import { Bus } from "../../bus.js"
+import { Catalog } from "../../catalog.js"
+import { Credential } from "../../credential.js"
+import { Integration } from "../../integration.js"
+import { ModalModels } from "../../modal/models.js"
+import { Model } from "../../model.js"
+import { Provider } from "../../provider.js"
+import type { PluginInternal } from "../internal.js"
+
+const providerID = Provider.ID.make("modal")
+
+export const ModalPlugin = define({
+  id: "opencode.provider.modal",
+  effect: Effect.fn(function* (ctx) {
+    const catalog = yield* Catalog.Service
+    const bus = yield* Bus.Service
+    const loading = Semaphore.makeUnsafe(1)
+    const loaded: {
+      baseURL?: string
+      models?: Map<Model.ID, Model.Info>
+    } = {}
+
+    const load = Effect.fn("ModalPlugin.load")(function* () {
+      const connection = yield* ctx.integration.connection.active("modal")
+      const credential = connection
+        ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
+        : undefined
+      const apiKey = credential?.type === "key" ? credential.key : process.env.MODAL_PROXY_TOKEN
+      const provider = yield* catalog.provider.get(providerID)
+      const baseURL = typeof provider?.settings?.baseURL === "string" ? provider.settings.baseURL : undefined
+      if (!apiKey || !baseURL) {
+        loaded.baseURL = undefined
+        loaded.models = undefined
+        return
+      }
+      loaded.baseURL = baseURL
+      const existing = (yield* catalog.model.all()).filter((model) => model.providerID === providerID)
+      loaded.models = yield* Effect.tryPromise({
+        try: () => ModalModels.get(baseURL, apiKey, existing),
+        catch: (cause) => cause,
+      }).pipe(
+        Effect.catch((cause) => Effect.logWarning("failed to sync Modal models", { cause }).pipe(Effect.as(undefined))),
+      )
+    })
+
+    yield* ctx.catalog.transform((evt) => {
+      const item = evt.provider.get(providerID)
+      if (!item) return
+      if (!loaded.models) return
+      for (const id of item.models.keys()) {
+        if (!loaded.models.has(Model.ID.make(id))) evt.model.remove(item.provider.id, id)
+      }
+      for (const [id, model] of loaded.models) {
+        evt.model.update(item.provider.id, id, (draft) => Object.assign(draft, structuredClone(model)))
+      }
+    })
+    const refresh = () => loading.withPermit(load().pipe(Effect.andThen(ctx.catalog.reload())))
+    yield* bus.subscribe(Credential.Event.Switched).pipe(
+      Stream.filter((event) => event.data.integrationID === Integration.ID.make("modal")),
+      Stream.runForEach(refresh),
+      Effect.forkScoped({ startImmediately: true }),
+    )
+    yield* refresh().pipe(Effect.forkScoped)
+  }),
+} satisfies PluginInternal.InternalPlugin)

packages/core/test/modal/models.test.ts

@@ -0,0 +1,137 @@
+import { expect, test } from "bun:test"
+import { ModalModels } from "@opencode-ai/core/modal/models"
+import { Model } from "@opencode-ai/core/model"
+import { Provider } from "@opencode-ai/core/provider"
+import { ProviderPlugins } from "@opencode-ai/core/plugin/provider"
+import { Money } from "@opencode-ai/schema/money"
+
+const providerID = Provider.ID.make("modal")
+
+test("modal plugin is registered", () => {
+  expect(ProviderPlugins.map((item) => item.id)).toContain("opencode.provider.modal")
+})
+
+function template(id: string, overrides: Partial<Model.Info> = {}) {
+  return Model.Info.make({
+    ...Model.Info.default(providerID, Model.ID.make(id)),
+    name: `${id} catalog`,
+    family: Model.Family.make("catalog-family"),
+    ...overrides,
+  })
+}
+
+test("maps live Modal models onto catalog templates", async () => {
+  const server = Bun.serve({
+    port: 0,
+    fetch: (request) => {
+      expect(request.headers.get("Authorization")).toBe("Bearer test-key")
+      expect(new URL(request.url).pathname).toBe("/v1/models")
+      return Response.json({
+        data: [
+          {
+            id: "live-model",
+            base_model_id: "base-model",
+            name: "Live Model",
+            input_modalities: ["text", "image"],
+            output_modalities: ["text"],
+            context_length: 128000,
+            max_output_length: 8192,
+            pricing: { prompt: "0.000001", completion: 0.000002, input_cache_read: "0.0000002" },
+            supported_sampling_parameters: ["temperature"],
+            supported_features: ["tools", "reasoning"],
+            reasoning_options: [{ type: "effort", values: ["low", "high", null] }],
+            interleaved: { field: "reasoning_content" },
+          },
+          {
+            id: "standalone",
+            context_length: 64000,
+          },
+          { id: "malformed", context_length: "huge" },
+        ],
+      })
+    },
+  })
+
+  try {
+    const base = template("base-model")
+    const stale = template("stale")
+    const models = await ModalModels.get(`${server.url.origin}/v1`, "test-key", [base, stale])
+
+    expect(models.has(Model.ID.make("stale"))).toBe(false)
+    expect(models.has(Model.ID.make("malformed"))).toBe(false)
+
+    const model = models.get(Model.ID.make("live-model"))
+    expect(model?.name).toBe("Live Model")
+    expect(model?.family).toBe(Model.Family.make("catalog-family"))
+    expect(model?.providerID).toBe(providerID)
+    expect(model?.modelID).toBe(Model.ID.make("live-model"))
+    expect(model?.package).toBe(Provider.aisdk("@ai-sdk/openai-compatible"))
+    expect(model?.settings).toMatchObject({ baseURL: `${server.url.origin}/v1` })
+    expect(model?.compatibility).toMatchObject({ reasoningField: "reasoning_content" })
+    expect(model?.capabilities).toMatchObject({ tools: true, input: ["text", "image"], output: ["text"] })
+    expect(model?.cost[0]?.input).toBe(Money.USDPerMillionTokens.make(1))
+    expect(model?.cost[0]?.output).toBe(Money.USDPerMillionTokens.make(2))
+    expect(Number(model?.cost[0]?.cache.read)).toBeCloseTo(0.2, 10)
+    expect(model?.cost[0]?.cache.write).toBe(Money.USDPerMillionTokens.zero)
+    expect(model?.limit).toMatchObject({ context: 128000, output: 8192 })
+    expect(model?.variants.map((variant) => variant.id)).toEqual([
+      Model.VariantID.make("low"),
+      Model.VariantID.make("high"),
+      Model.VariantID.make("none"),
+    ])
+    expect(model?.variants[0]?.settings).toMatchObject({ reasoningEffort: "low" })
+    expect(model?.status).toBe("active")
+
+    const fresh = models.get(Model.ID.make("standalone"))
+    expect(fresh?.name).toBe("standalone")
+    expect(fresh?.family).toBeUndefined()
+    expect(fresh?.capabilities).toMatchObject({ tools: true, input: ["text"], output: ["text"] })
+    expect(fresh?.variants).toEqual([])
+    expect(fresh?.limit).toMatchObject({ context: 64000, output: 0 })
+  } finally {
+    await server.stop(true)
+  }
+})
+
+test("keeps template cost and limits when the proxy omits them", async () => {
+  const server = Bun.serve({
+    port: 0,
+    fetch: () =>
+      Response.json({
+        data: [{ id: "sparse", hugging_face_id: "hf-base" }],
+      }),
+  })
+
+  try {
+    const base = template("hf-base", {
+      cost: [
+        {
+          input: Money.USDPerMillionTokens.make(5),
+          output: Money.USDPerMillionTokens.make(10),
+          cache: { read: Money.USDPerMillionTokens.make(1), write: Money.USDPerMillionTokens.make(2) },
+        },
+      ],
+      limit: { context: 1000, input: 500, output: 250 },
+    })
+    const models = await ModalModels.get(server.url.origin, "test-key", [base])
+    const model = models.get(Model.ID.make("sparse"))
+    expect(model?.name).toBe("hf-base catalog")
+    expect(model?.cost[0]).toMatchObject({ input: 5, output: 10, cache: { read: 1, write: 2 } })
+    expect(model?.limit).toMatchObject({ context: 1000, input: 500, output: 250 })
+  } finally {
+    await server.stop(true)
+  }
+})
+
+test("throws on proxy failure so the plugin can fail soft", async () => {
+  const server = Bun.serve({
+    port: 0,
+    fetch: () => new Response("nope", { status: 500 }),
+  })
+
+  try {
+    await expect(ModalModels.get(server.url.origin, "test-key", [])).rejects.toThrow()
+  } finally {
+    await server.stop(true)
+  }
+})