feat(provider)!: replace models.dev with a static model catalog (#40)

.opencode/agent/translator.md

@@ -565,7 +565,6 @@ AICTRL_DISABLE_AUTOCOMPACT
 AICTRL_DISABLE_AUTOUPDATE
 AICTRL_DISABLE_DEFAULT_PLUGINS
 AICTRL_DISABLE_FILETIME_CHECK
-AICTRL_DISABLE_MODELS_FETCH
 AICTRL_DISABLE_PRUNE
 AICTRL_ENABLE_EXA
 AICTRL_ENABLE_EXPERIMENTAL_MODELS
@@ -577,7 +576,6 @@ AICTRL_EXPERIMENTAL_OXFMT
 AICTRL_FAKE_VCS
 AICTRL_GIT_BASH_PATH
 AICTRL_MODEL
-AICTRL_MODELS_URL
 AICTRL_PERMISSION
 AICTRL_PORT
 PROJECT_ROOT

.prettierignore

@@ -1,2 +1,3 @@
 sst-env.d.ts
 packages/desktop/src/bindings.ts
+packages/cli/src/provider/catalog.json

CONTRIBUTING.md

@@ -26,8 +26,9 @@ Want to take on an issue? Leave a comment and a maintainer may assign it to you
 
 ## Adding New Providers
 
-New providers shouldn't require many if ANY code changes, but if you want to add support for a new provider first make a PR to:
-https://github.com/anomalyco/models.dev
+Built-in providers and models live in the static catalog at `packages/cli/src/provider/catalog.json` (one model per line, in the `Provider.Model` shape; `api.url` is left out for providers whose SDK knows its own endpoint, and each provider also carries `npm`/`api` defaults for models added through config). To add or update a built-in model, edit that file; providers that need custom loading also need an entry in `CUSTOM_LOADERS` in `packages/cli/src/provider/provider.ts`.
+
+Any other OpenAI-compatible provider does not need a code change: users define it under `provider` in `aictrl.json`.
 
 ## Developing Aictrl
 

EVENTS.md

@@ -268,7 +268,7 @@ For compatibility with sessions written before usage provenance was persisted, a
 **`context`** (context-window utilization):
 
 - `used` (number) — tokens occupying the model's context window this turn: `input + cache.read + cache.write`.
-- `limit` (number) — the model's total context-window size in tokens, sourced from the models.dev registry (`model.limit.context`).
+- `limit` (number) — the model's total context-window size in tokens, sourced from the provider registry: the built-in model catalog plus `provider` config (`model.limit.context`). With `aictrl run --attach`, the attached server's provider registry is used instead.
 - `ratio` (number) — `used / limit` (≥0; may exceed 1 if usage exceeds the model's registered limit). A value approaching or exceeding 1 signals context-exhaustion risk.
 - `null` — emitted when the model's context limit is not known (e.g. unregistered custom endpoint), or usage is missing.
 

nix/opencode.nix

@@ -5,7 +5,6 @@
   bun,
   sysctl,
   makeBinaryWrapper,
-  models-dev,
   ripgrep,
   installShellFiles,
   versionCheckHook,
@@ -21,7 +20,6 @@ stdenvNoCC.mkDerivation (finalAttrs: {
     bun
     installShellFiles
     makeBinaryWrapper
-    models-dev
     writableTmpDirAsHomeHook
   ];
 
@@ -33,8 +31,6 @@ stdenvNoCC.mkDerivation (finalAttrs: {
     runHook postConfigure
   '';
 
-  env.MODELS_DEV_API_JSON = "${models-dev}/dist/_api.json";
-  env.AICTRL_DISABLE_MODELS_FETCH = true;
   env.AICTRL_VERSION = finalAttrs.version;
   env.AICTRL_CHANNEL = "local";
 
@@ -80,7 +76,7 @@ stdenvNoCC.mkDerivation (finalAttrs: {
     writableTmpDirAsHomeHook
   ];
   doInstallCheck = true;
-  versionCheckKeepEnvironment = [ "HOME" "AICTRL_DISABLE_MODELS_FETCH" ];
+  versionCheckKeepEnvironment = [ "HOME" ];
   versionCheckProgramArg = "--version";
 
   passthru = {

packages/cli/.gitignore

@@ -2,4 +2,3 @@ research
 dist
 gen
 app.log
-src/provider/models-snapshot.ts

packages/cli/script/build.ts

@@ -18,28 +18,6 @@ const singleFlag = process.argv.includes("--single") || process.env.AICTRL_BUILD
 const baselineFlag = process.argv.includes("--baseline")
 const liteFlag = process.argv.includes("--lite")
 
-const modelsUrl = process.env.AICTRL_MODELS_URL || "https://models.dev"
-// Fetch and generate models.dev snapshot
-let modelsData = process.env.MODELS_DEV_API_JSON
-  ? await Bun.file(process.env.MODELS_DEV_API_JSON).text()
-  : await fetch(`${modelsUrl}/api.json`).then((x) => x.text())
-
-if (liteFlag) {
-  const allModels = JSON.parse(modelsData)
-  const liteProviders = ["anthropic", "openai", "google", "google-vertex", "openrouter", "zai", "aictrl"]
-  const filtered: Record<string, any> = {}
-  for (const key of liteProviders) {
-    if (allModels[key]) filtered[key] = allModels[key]
-  }
-  modelsData = JSON.stringify(filtered)
-}
-
-await Bun.write(
-  path.join(dir, "src/provider/models-snapshot.ts"),
-  `// Auto-generated by build.ts - do not edit\nexport const snapshot = ${modelsData} as const\n`,
-)
-console.log("Generated models-snapshot.ts")
-
 // Load migrations from migration directories
 const migrationDirs = (
   await fs.promises.readdir(path.join(dir, "migration"), {

packages/cli/src/cli/cmd/auth.ts

@@ -2,8 +2,8 @@ import { Auth } from "../../auth"
 import { cmd } from "./cmd"
 import * as prompts from "@clack/prompts"
 import { UI } from "../ui"
-import { ModelsDev } from "../../provider/models"
-import { map, pipe, sortBy, values } from "remeda"
+import { ModelCatalog } from "../../provider/catalog"
+import { map, pickBy, pipe, sortBy, values } from "remeda"
 import path from "path"
 import os from "os"
 import { Config } from "../../config/config"
@@ -163,7 +163,7 @@ async function handlePluginAuth(plugin: { auth: PluginAuth }, provider: string):
 
 /**
  * Build a deduplicated list of plugin-registered auth providers that are not
- * already present in models.dev, respecting enabled/disabled provider lists.
+ * already built into the model catalog, respecting enabled/disabled provider lists.
  * Pure function with no side effects; safe to test without mocking.
  */
 export function resolvePluginProviders(input: {
@@ -212,7 +212,7 @@ export const AuthListCommand = cmd({
     const displayPath = authPath.startsWith(homedir) ? authPath.replace(homedir, "~") : authPath
     prompts.intro(`Credentials ${UI.Style.TEXT_DIM}${displayPath}`)
     const results = Object.entries(await Auth.all())
-    const database = await ModelsDev.get()
+    const database = ModelCatalog.providers()
 
     for (const [providerID, result] of results) {
       const name = database[providerID]?.name || providerID
@@ -288,30 +288,21 @@ export const AuthLoginCommand = cmd({
           prompts.outro("Done")
           return
         }
-        await ModelsDev.refresh().catch(() => {})
-
         const config = await Config.get()
 
         const disabled = new Set(config.disabled_providers ?? [])
         const enabled = config.enabled_providers ? new Set(config.enabled_providers) : undefined
 
-        const providers = await ModelsDev.get().then((x) => {
-          const filtered: Record<string, (typeof x)[string]> = {}
-          for (const [key, value] of Object.entries(x)) {
-            if ((enabled ? enabled.has(key) : true) && !disabled.has(key)) {
-              filtered[key] = value
-            }
-          }
-          return filtered
-        })
+        const providers = pickBy(
+          ModelCatalog.providers(),
+          (_, id) => (enabled ? enabled.has(id) : true) && !disabled.has(id),
+        )
 
         const priority: Record<string, number> = {
-          aictrl: 0,
-          anthropic: 1,
-          openai: 2,
-          google: 3,
-          openrouter: 4,
-          vercel: 5,
+          anthropic: 0,
+          openai: 1,
+          openrouter: 2,
+          vercel: 3,
         }
         const pluginProviders = resolvePluginProviders({
           hooks: await Plugin.list(),
@@ -335,7 +326,6 @@ export const AuthLoginCommand = cmd({
                 label: x.name,
                 value: x.id,
                 hint: {
-                  aictrl: "recommended",
                   anthropic: "Claude Max or API key",
                   openai: "ChatGPT Plus/Pro or API key",
                 }[x.id],
@@ -392,10 +382,6 @@ export const AuthLoginCommand = cmd({
           )
         }
 
-        if (provider === "aictrl") {
-          prompts.log.info("Create an api key at https://aictrl.ai/auth")
-        }
-
         if (provider === "vercel") {
           prompts.log.info("You can create an api key at https://vercel.link/ai-gateway-token")
         }
@@ -433,7 +419,7 @@ export const AuthLogoutCommand = cmd({
       prompts.log.error("No credentials found")
       return
     }
-    const database = await ModelsDev.get()
+    const database = ModelCatalog.providers()
     const providerID = await prompts.select({
       message: "Select provider",
       options: credentials.map(([key, value]) => ({

packages/cli/src/cli/cmd/models.ts

@@ -1,7 +1,6 @@
 import type { Argv } from "yargs"
 import { Instance } from "../../project/instance"
 import { Provider } from "../../provider/provider"
-import { ModelsDev } from "../../provider/models"
 import { cmd } from "./cmd"
 import { UI } from "../ui"
 import { EOL } from "os"
@@ -20,17 +19,8 @@ export const ModelsCommand = cmd({
         describe: "use more verbose model output (includes metadata like costs)",
         type: "boolean",
       })
-      .option("refresh", {
-        describe: "refresh the models cache from models.dev",
-        type: "boolean",
-      })
   },
   handler: async (args) => {
-    if (args.refresh) {
-      await ModelsDev.refresh()
-      UI.println(UI.Style.TEXT_SUCCESS_BOLD + "Models cache refreshed" + UI.Style.TEXT_NORMAL)
-    }
-
     await Instance.provide({
       directory: process.cwd(),
       async fn() {
@@ -60,13 +50,7 @@ export const ModelsCommand = cmd({
           return
         }
 
-        const providerIDs = Object.keys(providers).sort((a, b) => {
-          const aIsOpencode = a.startsWith("aictrl")
-          const bIsOpencode = b.startsWith("aictrl")
-          if (aIsOpencode && !bIsOpencode) return -1
-          if (!aIsOpencode && bIsOpencode) return 1
-          return a.localeCompare(b)
-        })
+        const providerIDs = Object.keys(providers).sort((a, b) => a.localeCompare(b))
 
         for (const providerID of providerIDs) {
           printModels(providerID, args.verbose)

packages/cli/src/cli/cmd/run.ts

@@ -17,7 +17,7 @@ import { Filesystem } from "../../util/filesystem"
 import { createAictrlClient } from "@aictrl/sdk"
 import type { ToolPart } from "@aictrl/sdk/v2"
 import { Provider } from "../../provider/provider"
-import { ModelsDev } from "../../provider/models"
+import { lazy } from "../../util/lazy"
 import { Agent } from "../../agent/agent"
 import { PermissionNext } from "../../permission/next"
 import { Session } from "../../session"
@@ -120,9 +120,12 @@ export function buildContextWindow(
 export async function attachedContextLimit(
   providerID: string,
   modelID: string,
-  providers = ModelsDev.get(),
+  providers: Promise<Provider.Info[]>,
 ): Promise<number | null> {
-  return providers.then((providers) => providers[providerID]?.models[modelID]?.limit.context ?? null).catch(() => null)
+  return providers
+    .then((providers) => providers.find((provider) => provider.id === providerID))
+    .then((provider) => provider?.models[modelID]?.limit.context ?? null)
+    .catch(() => null)
 }
 
 type TerminalUsageInfo = Pick<MessageV2.Assistant, "finish" | "usageStatus" | "tokens">
@@ -542,6 +545,11 @@ export const RunCommand = cmd({
     let promptResult: Promise<any> = Promise.resolve()
 
     async function execute(sdk: any) {
+      // An attached server resolves models from its own provider registry, so ask it once per run.
+      const attachedProviders = lazy(
+        (): Promise<Provider.Info[]> => sdk.config.providers().then((res: any) => res.data?.providers ?? []),
+      )
+
       function tool(part: ToolPart) {
         try {
           if (part.tool === "bash") return bash(props<typeof BashTool>(part))
@@ -697,10 +705,10 @@ export const RunCommand = cmd({
                 // part of the prompt sent to the model and count against the context
                 // window, just billed at the cache-write rate. Excluding them
                 // undercounts utilization on the first turn of a conversation.
-                // limit comes from the model registry (models.dev). On lookup failure
-                // (or limit===0 for custom models) buildContextWindow returns null.
+                // limit comes from the provider registry (the built-in catalog plus config, or the attached server's).
+                // On lookup failure (or limit===0 for custom models) buildContextWindow returns null.
                 const contextLimit = await (args.attach
-                  ? attachedContextLimit(info.providerID, info.modelID)
+                  ? attachedContextLimit(info.providerID, info.modelID, attachedProviders())
                   : Provider.getModel(info.providerID, info.modelID)
                       .then((m) => m.limit.context)
                       .catch((e) => {

packages/cli/src/cli/error.ts

@@ -14,8 +14,10 @@ export function FormatError(input: unknown) {
       ...(Array.isArray(suggestions) && suggestions.length ? ["Did you mean: " + suggestions.join(", ")] : []),
       `Try: \`aictrl models\` to list available models`,
       `Or check your config (aictrl.json) provider/model names`,
+      `Providers that are not built in must be defined under "provider" in aictrl.json`,
     ].join("\n")
   }
+  if (Provider.BaseURLMissingError.isInstance(input)) return input.message
   if (Provider.InitError.isInstance(input)) {
     return `Failed to initialize provider "${input.data.providerID}". Check credentials and configuration.`
   }