@jacobbd/relay-ai 0.6.2 → 0.7.0

package/README.md

@@ -17,9 +17,9 @@
 
 📺 **Watch the Demos**
 
-| **Claude Code / Cowork / Desktop** | **Codex CLI & Desktop App** | **Gemini CLI** | **v0.4.1: UI & Antigravity** |
-|:---:|:---:|:---:|:---:|
-| [![Claude Demo](https://img.youtube.com/vi/IvsUPHLhX0o/mqdefault.jpg)](https://youtu.be/IvsUPHLhX0o) | [![Codex Demo](https://img.youtube.com/vi/42oiOB8IAu4/mqdefault.jpg)](https://youtu.be/42oiOB8IAu4) | [![Gemini Demo](https://img.youtube.com/vi/g7JKvqOHJl4/mqdefault.jpg)](https://www.youtube.com/watch?v=g7JKvqOHJl4) | [![UI & Antigravity Demo](https://img.youtube.com/vi/8vXJ0LfpdoY/mqdefault.jpg)](https://www.youtube.com/watch?v=8vXJ0LfpdoY) |
+| **Claude Code / Cowork / Desktop** | **Codex CLI & Desktop App** | **Gemini CLI** | **v0.4.1: UI & Antigravity** | **Server Gateway** |
+|:---:|:---:|:---:|:---:|:---:|
+| [![Claude Demo](https://img.youtube.com/vi/IvsUPHLhX0o/mqdefault.jpg)](https://youtu.be/IvsUPHLhX0o) | [![Codex Demo](https://img.youtube.com/vi/42oiOB8IAu4/mqdefault.jpg)](https://youtu.be/42oiOB8IAu4) | [![Gemini Demo](https://img.youtube.com/vi/g7JKvqOHJl4/mqdefault.jpg)](https://www.youtube.com/watch?v=g7JKvqOHJl4) | [![UI & Antigravity Demo](https://img.youtube.com/vi/8vXJ0LfpdoY/mqdefault.jpg)](https://www.youtube.com/watch?v=8vXJ0LfpdoY) | [![Server Gateway Demo](https://img.youtube.com/vi/4CEw0MQUE9I/mqdefault.jpg)](https://www.youtube.com/watch?v=4CEw0MQUE9I) |
 
 **relay-ai** is an interactive CLI — and now a **visual launcher** — that connects AI coding tools to any provider and runs local API gateways on your machine. It supports **Claude Code**, **Claude Desktop (Cowork + Code)**, the **OpenAI Codex CLI**, the **ChatGPT desktop app in Codex mode (macOS + Windows)**, **Google Gemini CLI**, and experimental **Antigravity CLI / IDE** support.
 
@@ -53,6 +53,7 @@ Pick your backend:
 | `relay-ai antigravity-ide` | Launch Antigravity IDE with Relay models, macOS ([warning + guide](docs/ANTIGRAVITY.md)) |
 | `relay-ai providers auth <id>` | Authenticate an OAuth provider (GitHub Copilot, xAI, OpenAI) |
 | `relay-ai --ai` | Full agent reference for scripts and alef-agent ([guide](docs/AI-AGENTS.md)) |
+| `@jacobbd/relay-ai/core` | Embed Relay AI in a Node app in-process — no CLI/UI/server ([guide](docs/CORE.md)) |
 
 ## Features
 
@@ -406,6 +407,31 @@ export ANTHROPIC_API_KEY="anything"
 unset CLAUDE_CODE_USE_VERTEX ANTHROPIC_VERTEX_PROJECT_ID CLOUD_ML_REGION
 ```
 
+## Embedded usage (`@jacobbd/relay-ai/core`)
+
+Node.js applications can embed Relay AI in-process — no CLI, UI, or server — to list the configured model catalog and obtain ready [Vercel AI SDK](https://sdk.vercel.ai) `LanguageModel` instances:
+
+```ts
+import { listRelayModels, createRelayModel, isRelayCoreError } from '@jacobbd/relay-ai/core';
+import { streamText } from 'ai';
+
+const models = listRelayModels();          // credential-free catalog
+const model = await createRelayModel(models[0].routeId);
+const result = await streamText({ model, prompt: 'Hello!' });
+```
+
+**Route ids** are unconditionally scoped: `` `${providerId}::${modelId}` `` (split on the first `::`; model ids may contain `/` and `:`, e.g. `openrouter::vendor/model:free`). Because they are never bare, a route id persisted by your app stays valid even when a second provider later exposes the same model id.
+
+**Ownership boundary:** Relay AI keeps sole ownership of provider registration, credentials, the OS keyring, and OAuth login/refresh. `createRelayModel()` resolves credentials (refreshing expiring OAuth tokens) at call time and returns only the SDK model — your app never receives or stores credential material. Re-authentication always happens through Relay (`relay-ai ui`), never through the consumer.
+
+**No server required:** the Core API reads the same registry and config as the CLI (`~/.relay-ai`, overridable with `RELAY_AI_HOME`) and never starts a server, opens a browser, or writes to disk. `createRelayModel()` re-reads state on every call, so provider changes take effect without restarting your app. Errors are thrown as `RelayCoreError` with a machine-readable `code` (see `isRelayCoreError`) and never contain credential material.
+
+**Schema compatibility:** Core supports registry schema v1. A registry written by a newer Relay version fails fast with `UNSUPPORTED_REGISTRY_VERSION` — upgrade relay-ai rather than downgrading the file.
+
+**Capability detection:** `capabilities.tools` and `capabilities.vision` are always `'unknown'` today — the cached model catalog carries no tools/vision metadata, and Core deliberately never guesses from a model's name. Don't filter on `=== true` for these two fields; use `capabilities.reasoning` (`'none' | 'fixed' | 'adjustable' | 'unknown'`), which is derived from real provider metadata.
+
+**Full guide:** prerequisites, the complete API and error-code reference, and troubleshooting — see **[docs/CORE.md](docs/CORE.md)**.
+
 ## Antigravity CLI, app, and IDE support
 
 Relay AI can launch the Antigravity CLI, standalone Antigravity app, and Antigravity IDE through a local Cloud Code gateway. This lets Antigravity's native model picker show Relay models from your configured providers.
@@ -455,8 +481,9 @@ Launch **Claude Desktop** (macOS or Windows) with registry providers:
 relay-ai claude-app
 ```
 
-This command automates the "Third-Party Inference" (Developer Mode) setup. It temporarily configures Claude Desktop to point at a local gateway, launches the app, and routes traffic to your chosen provider.
+This command automates the "Third-Party Inference" (Developer Mode) setup. It temporarily configures Claude Desktop to point at a local gateway, launches the app, and exposes the selected model followed by your available saved favorites (up to 20 models total).
 
+- **Selected + favorites:** The selected model is listed first; duplicates and unavailable favorites are skipped. Claude Desktop ultimately controls which discovered model it activates initially.
 - **Keep the terminal open:** The proxy runs in the foreground.
 - **Ctrl+C to restore:** When you're done, press `Ctrl+C` in the terminal to automatically restore Claude Desktop to its normal Anthropic cloud mode.
 - **Cleanup:** If the terminal crashes, run `relay-ai claude-app --restore`.

package/dist/{chunk-I3I5LKZI.js → chunk-IYYLLN5T.js}

@@ -11,7 +11,7 @@ import { join } from "path";
 // package.json
 var package_default = {
   name: "@jacobbd/relay-ai",
-  version: "0.6.2",
+  version: "0.7.0",
   publishConfig: {
     access: "public"
   },
@@ -46,7 +46,7 @@ var package_default = {
     node: ">=18"
   },
   scripts: {
-    build: "tsup && mkdir -p dist/ui/public && cp -r src/ui/public/. dist/ui/public/",
+    build: "tsup && tsup --config tsup.core.config.ts && node scripts/copy-ui-assets.mjs",
     dev: "tsup --watch",
     test: "vitest run",
     "test:watch": "vitest",
@@ -100,6 +100,14 @@ var package_default = {
   },
   overrides: {
     ws: "^8.21.0"
+  },
+  exports: {
+    "./core": {
+      types: "./dist/core/index.d.ts",
+      import: "./dist/core/index.js",
+      default: "./dist/core/index.js"
+    },
+    "./package.json": "./package.json"
   }
 };
 
@@ -139,6 +147,14 @@ var CONFLICTING_ENV_VARS = [
   "ANTHROPIC_DEFAULT_SONNET_MODEL",
   "ANTHROPIC_DEFAULT_HAIKU_MODEL"
 ];
+var PARENT_SESSION_ENV_VARS = [
+  "CLAUDECODE",
+  "CLAUDE_CODE_CHILD_SESSION",
+  "CLAUDE_CODE_SESSION_ID",
+  "CLAUDE_CODE_HOST_SESSION_ID",
+  "CLAUDE_CODE_ENTRYPOINT",
+  "CLAUDE_PID"
+];
 var OPENCODE_CACHE_PATH = join(homedir(), ".cache", "opencode", "models.json");
 var MAX_MODEL_CATALOG = 20;
 var VERTEX_ANTHROPIC_NPM = "@ai-sdk/google-vertex/anthropic";
@@ -3045,6 +3061,9 @@ function buildChildEnv(baseUrl, model, apiKey, proxyPort, contextWindow, enableG
   for (const name of CONFLICTING_ENV_VARS) {
     delete env[name];
   }
+  for (const name of PARENT_SESSION_ENV_VARS) {
+    delete env[name];
+  }
   env["ANTHROPIC_BASE_URL"] = proxyPort ? `http://127.0.0.1:${proxyPort}` : baseUrl;
   env["ANTHROPIC_API_KEY"] = apiKey;
   const bareModel = stripOneMContextSuffix(model);
@@ -3616,7 +3635,7 @@ function parseRegistry(raw) {
   if (typeof data.pricingCacheAt === "string") registry.pricingCacheAt = data.pricingCacheAt;
   return registry;
 }
-function loadRegistry(path = getProvidersPath()) {
+function loadRegistry(path = getProvidersPath(), { persist = true } = {}) {
   if (!existsSync6(path)) {
     return { schemaVersion: REGISTRY_SCHEMA_VERSION, providers: [] };
   }
@@ -3627,7 +3646,7 @@ function loadRegistry(path = getProvidersPath()) {
     if (migrateOAuthOpenAiProvider(registry)) migrated = true;
     if (migrateOAuthXaiProvider(registry)) migrated = true;
     if (migrateAlibabaDashScopeChinaLabel(registry)) migrated = true;
-    if (migrated) {
+    if (migrated && persist) {
       try {
         saveRegistry(registry, path);
       } catch {
@@ -4243,7 +4262,7 @@ function maskGatewayModelId(aliasId) {
 // src/server/models.ts
 var CREATED_AT_ISO = "2025-01-01T00:00:00Z";
 var CREATED_AT_UNIX = 1735689600;
-function formatAnthropicModelEntry(id, displayName, contextWindow) {
+function formatAnthropicModelEntry(id, displayName, contextWindow, options) {
   const maxInput = resolveContextWindow(id, contextWindow);
   return {
     id,
@@ -4251,12 +4270,18 @@ function formatAnthropicModelEntry(id, displayName, contextWindow) {
     display_name: displayName,
     created_at: CREATED_AT_ISO,
     context_window: maxInput,
-    max_input_tokens: maxInput
+    max_input_tokens: maxInput,
+    ...options?.supportsOneM !== void 0 ? { supports_1m: options.supportsOneM } : {}
   };
 }
 function formatAnthropicModelList(entries) {
   return {
-    data: entries.map((entry) => formatAnthropicModelEntry(entry.id, entry.name, entry.contextWindow)),
+    data: entries.map((entry) => formatAnthropicModelEntry(
+      entry.id,
+      entry.name,
+      entry.contextWindow,
+      { supportsOneM: entry.supportsOneM }
+    )),
     has_more: false,
     first_id: entries[0]?.id ?? null,
     last_id: entries.at(-1)?.id ?? null
@@ -4282,22 +4307,28 @@ function openAiExposedId(model, collisions) {
   return collisions.has(model.id) ? `${gatewayProviderId(model)}/${model.id}` : model.id;
 }
 function exposedGatewayAliasId(model, opts) {
-  const alias = gatewayAliasId(model);
-  return opts?.maskGatewayIds ? maskGatewayModelId(alias) : alias;
+  const singleOneM = usesSingleOneMEntry(model, opts);
+  const alias = gatewayAliasId(singleOneM ? { ...model, id: stripOneMContextSuffix(model.id) } : model);
+  const exposed = opts?.maskGatewayIds ? maskGatewayModelId(alias) : alias;
+  return singleOneM ? `${stripOneMContextSuffix(exposed)}[1m]` : exposed;
 }
 function gatewayDisplayName(model, opts) {
-  if (!opts?.maskGatewayIds) return model.name;
-  return `${model.name} (${gatewayProviderLabel(model)})`;
+  const name = opts?.maskGatewayIds ? `${model.name} (${gatewayProviderLabel(model)})` : model.name;
+  return usesSingleOneMEntry(model, opts) && !/\b1m$/i.test(name) ? `${name} 1M` : name;
 }
 function formatGatewayAnthropicModels(models, opts) {
   return formatAnthropicModelList(
     models.map((model) => ({
       id: exposedGatewayAliasId(model, opts),
       name: gatewayDisplayName(model, opts),
-      contextWindow: model.contextWindow
+      contextWindow: model.contextWindow,
+      supportsOneM: usesSingleOneMEntry(model, opts) ? false : void 0
     }))
   );
 }
+function usesSingleOneMEntry(model, opts) {
+  return opts?.longContextDisplay === "single-1m" && resolveContextWindow(stripOneMContextSuffix(model.id), model.contextWindow) >= 1e6;
+}
 function createGatewayModelCatalog(models, opts) {
   const byId = /* @__PURE__ */ new Map();
   const collisions = openAiIdCollisions(models);
@@ -4307,8 +4338,23 @@ function createGatewayModelCatalog(models, opts) {
     if (scopedId !== model.id) byId.set(scopedId, model);
     const alias = exposedGatewayAliasId(model, opts);
     if (alias !== model.id) byId.set(alias, model);
+    const singleOneM = usesSingleOneMEntry(model, opts);
+    if (singleOneM) {
+      const bareModel = { ...model, id: stripOneMContextSuffix(model.id) };
+      const rawBareAlias = gatewayAliasId(bareModel);
+      const exposedBareAlias = opts?.maskGatewayIds ? maskGatewayModelId(rawBareAlias) : rawBareAlias;
+      for (const compatibleId of [
+        stripOneMContextSuffix(model.id),
+        rawBareAlias,
+        `${rawBareAlias}[1m]`,
+        exposedBareAlias,
+        `${exposedBareAlias}[1m]`
+      ]) {
+        byId.set(compatibleId, model);
+      }
+    }
     if (opts?.maskGatewayIds) {
-      const rawAlias = gatewayAliasId(model);
+      const rawAlias = gatewayAliasId(singleOneM ? { ...model, id: stripOneMContextSuffix(model.id) } : model);
       if (rawAlias !== alias) byId.set(rawAlias, model);
     }
   }
@@ -10679,6 +10725,8 @@ function winLocalAppData2() {
 function winClaudeExeCandidates() {
   const local = winLocalAppData2();
   const bases = [
+    // Squirrel install folder used by the Anthropic Claude desktop installer.
+    join12(local, "AnthropicClaude"),
     join12(local, "Programs", "Claude"),
     join12(local, "Claude")
   ];
@@ -11063,4 +11111,4 @@ export {
   supportsClaudeTransparentMode,
   buildHttpProxyRoutes
 };
-//# sourceMappingURL=chunk-I3I5LKZI.js.map
+//# sourceMappingURL=chunk-IYYLLN5T.js.map