@myvillage/cli 1.80.0 → 1.81.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@myvillage/cli",
-  "version": "1.80.0",
+  "version": "1.81.0",
   "description": "MyVillageOS CLI for community developers",
   "type": "module",
   "bin": {

package/src/agent-runtime/loop.js

@@ -64,7 +64,9 @@ export async function agentLoop(agentName, { signal }) {
     }
     const apiKey = process.env.SELF_HOSTED_API_KEY || globalConfig.selfHostedApiKey || 'not-needed';
     const openai = createOpenAI({ apiKey, baseURL });
-    model = openai(modelId);
+    // Most OpenAI-compatible servers (vLLM, Ollama, LM Studio, Baseten) serve
+    // /chat/completions but not the Responses API that openai(id) defaults to.
+    model = openai.chat(modelId);
   } else if (provider === 'myvillage') {
     // MyVillage-hosted models via the AI Gateway. OpenAI-compatible; the villager's
     // OAuth token is the credential, and MVTs are deducted server-side by usage.
@@ -74,8 +76,32 @@ export async function agentLoop(agentName, { signal }) {
       throw new Error('Not logged in. Run `myvillage login` to use MyVillage-hosted models.');
     }
     const { apiBaseUrl } = getConfig();
-    const openai = createOpenAI({ apiKey: token, baseURL: `${apiBaseUrl}/ai` });
-    model = openai(modelId);
+    // The model client is built once, but OAuth tokens rotate (~1h). A static
+    // bearer would 401 on every gateway call after expiry for long-lived agents,
+    // so inject the current token per request — refreshing first if it expired.
+    // Tell the gateway which agent is spending MVTs, so the server can attribute the
+    // MVT transaction (e.g. "AI chat · Vee 1.3 · agent: research-bot"). Header values
+    // must be safe ASCII — undici rejects CR/LF and non-latin1 — so strip aggressively.
+    const agentHeader = String(config.display_name || agentName || '')
+      .replace(/[^\x20-\x7e]+/g, ' ')
+      .replace(/\s+/g, ' ')
+      .trim()
+      .slice(0, 64);
+    const gatewayFetch = async (url, init = {}) => {
+      if (isTokenExpired()) {
+        try { await refreshAccessToken(); } catch { /* fall through with the existing token */ }
+      }
+      const current = getAccessToken();
+      const headers = new Headers(init.headers);
+      if (current) headers.set('authorization', `Bearer ${current}`);
+      if (agentHeader) headers.set('x-myvillage-agent', agentHeader);
+      return fetch(url, { ...init, headers });
+    };
+    const openai = createOpenAI({ apiKey: token, baseURL: `${apiBaseUrl}/ai`, fetch: gatewayFetch });
+    // The AI Gateway implements the OpenAI Chat Completions API (/chat/completions).
+    // @ai-sdk/openai's default `openai(id)` targets the newer Responses API
+    // (/responses), which the gateway does not serve — so pin to chat explicitly.
+    model = openai.chat(modelId);
   } else {
     const apiKey = process.env.ANTHROPIC_API_KEY || globalConfig.anthropicApiKey;
     if (!apiKey) {