@agentproto/llm-endpoint 0.9.0 → 0.10.0

package/dist/cli.mjs

@@ -1895,6 +1895,29 @@ function resolveNebiusBaseUrl(raw = process.env.NEBIUS_BASE_URL) {
   const value = raw?.trim() || CONFIGURABLE_PROVIDERS.nebius.defaultBaseUrl;
   return parseConfigurableUpstreamUrl(value, "nebius");
 }
+var DEVICE_PROVIDER_RE = /^(.+)@([^@/]+)$/;
+function parseDeviceProvider(provider) {
+  const m = DEVICE_PROVIDER_RE.exec(provider);
+  return m ? { endpointId: m[1], device: m[2] } : null;
+}
+function resolveDeviceProviderSpec(device) {
+  const unavailableMessage = () => `"@${device}" device-inference routing requires this llm-endpoint sidecar to be started by an agentproto daemon (LLM_ENDPOINT_DAEMON_URL is unset) \u2014 run it via \`agentproto serve\` with features.llmEndpoint on, not as a bare standalone process.`;
+  return {
+    keyRequired: true,
+    apiKeyEnv: "LLM_ENDPOINT_DAEMON_TOKEN",
+    resolveUpstream: () => {
+      const daemonUrl = process.env.LLM_ENDPOINT_DAEMON_URL?.trim();
+      if (!daemonUrl) return null;
+      const base = parseUpstreamUrl(daemonUrl);
+      if (!base) return null;
+      return {
+        ...base,
+        pathPrefix: `${base.pathPrefix}/devices/${encodeURIComponent(device)}/exec-stream/device-inference/v1`
+      };
+    },
+    unavailableMessage
+  };
+}
 function getConfigurableProviderSpec(provider) {
   const staticSpec = CONFIGURABLE_PROVIDERS[provider];
   if (staticSpec) {
@@ -1916,6 +1939,8 @@ function getConfigurableProviderSpec(provider) {
       unavailableMessage: () => `"${provider}" endpoint is misconfigured (invalid baseUrl).`
     };
   }
+  const deviceRoute = parseDeviceProvider(provider);
+  if (deviceRoute) return resolveDeviceProviderSpec(deviceRoute.device);
   return void 0;
 }
 function applyDefaultRequestFields(payload, defaults, skipKeys) {
@@ -2017,7 +2042,7 @@ async function probeFileEndpointModels(endpoint) {
   return result;
 }
 function isKnownProvider(provider) {
-  return KNOWN_PROVIDERS.has(provider) || getConfiguredEndpoints().some((e) => e.id === provider);
+  return KNOWN_PROVIDERS.has(provider) || getConfiguredEndpoints().some((e) => e.id === provider) || DEVICE_PROVIDER_RE.test(provider);
 }
 function applyProviderOverride(target, providerOverride) {
   const route = { provider: target.provider, model: target.model };
@@ -2034,7 +2059,7 @@ function parseAnyTransparentModel(model) {
   const slashIdx = model.indexOf("/");
   if (slashIdx <= 0 || slashIdx === model.length - 1) return null;
   const provider = model.slice(0, slashIdx);
-  if (!getConfiguredEndpoints().some((e) => e.id === provider)) return null;
+  if (!getConfiguredEndpoints().some((e) => e.id === provider) && !DEVICE_PROVIDER_RE.test(provider)) return null;
   return { provider, model: model.slice(slashIdx + 1) };
 }
 function resolveModelRoute(payload, ctx, localPacks = getLocalPacks()) {
@@ -2506,6 +2531,8 @@ function handleChatCompletionsRequest(req, res, opts) {
         return;
       }
       payload.model = resolvedTarget.model;
+      const deviceRoute = parseDeviceProvider(resolvedTarget.provider);
+      if (deviceRoute) payload.model = `${deviceRoute.endpointId}/${resolvedTarget.model}`;
       trimTools(payload, {
         provider: resolvedTarget.provider,
         queryTools: opts.queryTools,
@@ -2545,7 +2572,11 @@ function handleChatCompletionsRequest(req, res, opts) {
         method: "POST",
         headers: {
           "Content-Type": "application/json",
-          ...targetApiKey ? { "Authorization": `Bearer ${targetApiKey}` } : {}
+          ...targetApiKey ? { "Authorization": `Bearer ${targetApiKey}` } : {},
+          // See the /v1/messages handler's identical header — the daemon's
+          // /devices/:id/exec-stream 400s without it, always POST outer verb
+          // notwithstanding.
+          ...deviceRoute ? { "x-agentproto-forward-method": "POST" } : {}
         }
       };
       const proxyReq = sendUpstreamRequest(protocol, options, (proxyRes) => {
@@ -3382,6 +3413,8 @@ var server = createServer((req, res) => {
       let cred;
       let headers = { "Content-Type": "application/json" };
       payload.model = resolvedTarget.model;
+      const deviceRoute = parseDeviceProvider(resolvedTarget.provider);
+      if (deviceRoute) payload.model = `${deviceRoute.endpointId}/${resolvedTarget.model}`;
       trimTools(payload, {
         provider: resolvedTarget.provider,
         queryTools,
@@ -3407,6 +3440,7 @@ var server = createServer((req, res) => {
         cred = await resolveUpstreamCredential(resolvedTarget.provider);
         targetApiKey = cred?.value ?? "";
         if (cred && cred.value) Object.assign(headers, buildUpstreamAuthHeaders(resolvedTarget.provider, cred));
+        if (deviceRoute) headers["x-agentproto-forward-method"] = "POST";
         const clientThinkingEnabled = isRecord(payload.thinking) && payload.thinking.type === "enabled";
         adaptAnthropicToOpenAI(payload);
         applyDefaultRequestFields(