fix(chat ui): fixed llm inference chat with openclaw to work with new LiteLLM gateway (#13)

Author: nv-kasikritc

brev/welcome-ui/app.js

@@ -116,6 +116,10 @@
   let sandboxUrl = null;
   let installTriggered = false;
   let pollTimer = null;
+  let keyInjected = false;
+  let injectInFlight = false;
+  let injectTimer = null;
+  let lastSubmittedKey = "";
 
   function stopPolling() {
     if (pollTimer) {
@@ -124,12 +128,38 @@
     }
   }
 
+  async function submitKeyForInjection(key) {
+    if (key === lastSubmittedKey) return;
+    lastSubmittedKey = key;
+    keyInjected = false;
+    injectInFlight = true;
+    updateButtonState();
+    try {
+      await fetch("/api/inject-key", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ key }),
+      });
+    } catch {}
+    injectInFlight = false;
+    if (!pollTimer && sandboxReady) startPolling();
+  }
+
+  function onApiKeyInput() {
+    updateButtonState();
+    const key = apiKeyInput.value.trim();
+    if (!isApiKeyValid()) return;
+    if (injectTimer) clearTimeout(injectTimer);
+    injectTimer = setTimeout(() => submitKeyForInjection(key), 300);
+  }
+
   /**
-   * Four-state CTA button:
-   *  1. API empty  + tasks running   -> "Waiting for API key…"  (disabled)
-   *  2. API valid  + tasks running   -> "Provisioning Sandbox…" (disabled, spinner)
-   *  3. API empty  + tasks complete  -> "Waiting for API key…"  (disabled)
-   *  4. API valid  + tasks complete  -> "Open NemoClaw"          (enabled)
+   * Five-state CTA button:
+   *  1. API empty  + tasks running        -> "Waiting for API key…"      (disabled)
+   *  2. API valid  + tasks running        -> "Provisioning Sandbox…"     (disabled, spinner)
+   *  3. API empty  + tasks complete       -> "Waiting for API key…"      (disabled)
+   *  4. API valid  + sandbox ready + !key -> "Configuring API key…"      (disabled, spinner)
+   *  5. API valid  + sandbox ready + key  -> "Open NemoClaw"             (enabled)
    */
   function updateButtonState() {
     const keyValid = isApiKeyValid();
@@ -148,20 +178,26 @@
     }
 
     // Console "ready" line
-    if (sandboxReady && keyValid) {
+    if (sandboxReady && keyValid && keyInjected) {
       logReady.hidden = false;
       logReady.querySelector(".console__icon").textContent = CHECK_CHAR;
       logReady.querySelector(".console__icon").className = "console__icon console__icon--done";
     } else {
       logReady.hidden = true;
     }
 
-    if (sandboxReady && keyValid) {
+    if (sandboxReady && keyValid && keyInjected) {
       btnLaunch.disabled = false;
       btnLaunch.classList.add("btn--ready");
       btnSpinner.hidden = true;
       btnSpinner.style.display = "none";
       btnLaunchLabel.textContent = "Open NemoClaw";
+    } else if (sandboxReady && keyValid && !keyInjected) {
+      btnLaunch.disabled = true;
+      btnLaunch.classList.remove("btn--ready");
+      btnSpinner.hidden = false;
+      btnSpinner.style.display = "";
+      btnLaunchLabel.textContent = "Configuring API key\u2026";
     } else if (!sandboxReady && keyValid) {
       btnLaunch.disabled = true;
       btnLaunch.classList.remove("btn--ready");
@@ -229,19 +265,28 @@
         const res = await fetch("/api/sandbox-status");
         const data = await res.json();
 
+        if (!injectInFlight) {
+          keyInjected = !!data.key_injected;
+        }
+
         if (data.status === "running") {
-          stopPolling();
           sandboxReady = true;
           sandboxUrl = data.url || null;
 
           setLogIcon(logGatewayIcon, "done");
           logGateway.querySelector(".console__text").textContent =
             "OpenClaw agent gateway online.";
+
+          if (keyInjected) {
+            stopPolling();
+          }
           updateButtonState();
         } else if (data.status === "error") {
           stopPolling();
           installTriggered = false;
           showError(data.error || "Sandbox creation failed");
+        } else {
+          updateButtonState();
         }
       } catch {
         // transient fetch error, keep polling
@@ -250,7 +295,7 @@
   }
 
   function openOpenClaw() {
-    if (!sandboxReady || !isApiKeyValid() || !sandboxUrl) return;
+    if (!sandboxReady || !isApiKeyValid() || !keyInjected || !sandboxUrl) return;
 
     const apiKey = apiKeyInput.value.trim();
     const url = new URL(sandboxUrl);
@@ -262,6 +307,8 @@
     sandboxReady = false;
     sandboxUrl = null;
     installTriggered = false;
+    keyInjected = false;
+    lastSubmittedKey = "";
     stopPolling();
 
     setLogIcon(logSandboxIcon, null);
@@ -277,7 +324,7 @@
     triggerInstall();
   }
 
-  apiKeyInput.addEventListener("input", updateButtonState);
+  apiKeyInput.addEventListener("input", onApiKeyInput);
   btnLaunch.addEventListener("click", openOpenClaw);
   btnRetry.addEventListener("click", resetInstall);
 
@@ -288,6 +335,10 @@
       const res = await fetch("/api/sandbox-status");
       const data = await res.json();
 
+      if (data.key_injected) {
+        keyInjected = true;
+      }
+
       if (data.status === "running" && data.url) {
         sandboxReady = true;
         sandboxUrl = data.url;
@@ -302,6 +353,9 @@
         updateButtonState();
 
         showOverlay(overlayInstall);
+        if (!keyInjected) {
+          startPolling();
+        }
       } else if (data.status === "creating") {
         installTriggered = true;
 

brev/welcome-ui/index.html

@@ -230,6 +230,6 @@ <h4 class="instructions-section__title">4. Manage policies with the TUI</h4>
     NemoClaw Sandbox &middot; NVIDIA
   </footer>
 
-  <script src="app.js?v=11"></script>
+  <script src="app.js?v=12"></script>
 </body>
 </html>

brev/welcome-ui/server.py

@@ -5,6 +5,7 @@
 
 """NemoClaw Welcome UI — HTTP server with sandbox lifecycle APIs."""
 
+import hashlib
 import http.client
 import http.server
 import json
@@ -43,6 +44,66 @@
     "error": None,
 }
 
+_inject_key_lock = threading.Lock()
+_inject_key_state = {
+    "status": "idle",  # idle | injecting | done | error
+    "error": None,
+    "key_hash": None,
+}
+
+
+def _hash_key(key: str) -> str:
+    return hashlib.sha256(key.encode()).hexdigest()
+
+
+def _inject_log(msg: str) -> None:
+    ts = time.strftime("%H:%M:%S")
+    sys.stderr.write(f"[inject-key {ts}] {msg}\n")
+    sys.stderr.flush()
+
+
+def _run_inject_key(key: str, key_hash: str) -> None:
+    """Background thread: update the NemoClaw provider credential."""
+    _inject_log(f"step 1/3: received key (hash={key_hash[:12]}…)")
+    cmd = [
+        "nemoclaw", "provider", "update", "nvidia-inference",
+        "--type", "openai",
+        "--credential", f"OPENAI_API_KEY={key}",
+        "--config", "OPENAI_BASE_URL=https://inference-api.nvidia.com/v1",
+    ]
+    _inject_log(f"step 2/3: running nemoclaw provider update nvidia-inference …")
+    try:
+        t0 = time.time()
+        result = subprocess.run(
+            cmd, capture_output=True, text=True, timeout=120,
+        )
+        elapsed = time.time() - t0
+        _inject_log(f"         CLI exited {result.returncode} in {elapsed:.1f}s")
+        if result.stdout.strip():
+            _inject_log(f"         stdout: {result.stdout.strip()}")
+        if result.stderr.strip():
+            _inject_log(f"         stderr: {result.stderr.strip()}")
+
+        if result.returncode != 0:
+            err = (result.stderr or result.stdout or "unknown error").strip()
+            _inject_log(f"step 3/3: FAILED — {err}")
+            with _inject_key_lock:
+                _inject_key_state["status"] = "error"
+                _inject_key_state["error"] = err
+            return
+
+        _inject_log(f"step 3/3: SUCCESS — provider nvidia-inference updated")
+        with _inject_key_lock:
+            _inject_key_state["status"] = "done"
+            _inject_key_state["error"] = None
+            _inject_key_state["key_hash"] = key_hash
+
+    except Exception as exc:
+        _inject_log(f"step 3/3: EXCEPTION — {exc}")
+        with _inject_key_lock:
+            _inject_key_state["status"] = "error"
+            _inject_key_state["error"] = str(exc)
+
 
 def _sandbox_ready() -> bool:
     with _sandbox_lock:
@@ -391,6 +452,8 @@ def _route(self):
             return self._handle_install_openclaw()
         if path == "/api/policy-sync" and self.command == "POST":
             return self._handle_policy_sync()
+        if path == "/api/inject-key" and self.command == "POST":
+            return self._handle_inject_key()
 
         if _sandbox_ready():
             return self._proxy_to_sandbox()
@@ -551,6 +614,50 @@ def _handle_policy_sync(self):
         _log(f"── responding {status}: {json.dumps(result)}")
         return self._json_response(status, result)
 
+    # -- POST /api/inject-key -------------------------------------------
+
+    def _handle_inject_key(self):
+        """Asynchronously update the NemoClaw provider credential.
+
+        Returns immediately (202) and runs the slow CLI command in a
+        background thread.  The frontend polls /api/sandbox-status to
+        learn when injection is complete.
+        """
+        content_length = int(self.headers.get("Content-Length", 0))
+        if content_length == 0:
+            return self._json_response(400, {"ok": False, "error": "empty body"})
+        raw = self.rfile.read(content_length).decode("utf-8", errors="replace")
+        try:
+            data = json.loads(raw)
+        except json.JSONDecodeError:
+            return self._json_response(400, {"ok": False, "error": "invalid JSON"})
+
+        key = data.get("key", "").strip()
+        if not key:
+            return self._json_response(400, {"ok": False, "error": "missing key"})
+
+        key_hash = _hash_key(key)
+
+        with _inject_key_lock:
+            if (_inject_key_state["status"] == "done"
+                    and _inject_key_state["key_hash"] == key_hash):
+                return self._json_response(200, {"ok": True, "already": True})
+
+            if (_inject_key_state["status"] == "injecting"
+                    and _inject_key_state["key_hash"] == key_hash):
+                return self._json_response(202, {"ok": True, "started": True})
+
+            _inject_key_state["status"] = "injecting"
+            _inject_key_state["error"] = None
+            _inject_key_state["key_hash"] = key_hash
+
+        thread = threading.Thread(
+            target=_run_inject_key, args=(key, key_hash), daemon=True,
+        )
+        thread.start()
+
+        return self._json_response(202, {"ok": True, "started": True})
+
     # -- GET /api/sandbox-status ----------------------------------------
 
     def _handle_sandbox_status(self):
@@ -568,10 +675,16 @@ def _handle_sandbox_status(self):
             state["status"] = "running"
             state["url"] = url
 
+        with _inject_key_lock:
+            key_injected = _inject_key_state["status"] == "done"
+            key_inject_error = _inject_key_state.get("error")
+
         return self._json_response(200, {
             "status": state["status"],
             "url": state.get("url"),
             "error": state.get("error"),
+            "key_injected": key_injected,
+            "key_inject_error": key_inject_error,
         })
 
     # -- GET /api/connection-details ------------------------------------

sandboxes/nemoclaw/nemoclaw-start.sh

@@ -69,6 +69,7 @@ fi
 # Onboard and start the gateway
 # --------------------------------------------------------------------------
 export NVIDIA_API_KEY="${NVIDIA_INFERENCE_API_KEY:- }"
+_ONBOARD_KEY="${NVIDIA_INFERENCE_API_KEY:-not-used}"
 openclaw onboard \
   --non-interactive \
   --accept-risk \
@@ -79,7 +80,7 @@ openclaw onboard \
   --auth-choice custom-api-key \
   --custom-base-url "https://inference.local/v1" \
   --custom-model-id "aws/anthropic/bedrock-claude-opus-4-6" \
-  --custom-api-key "not-used" \
+  --custom-api-key "$_ONBOARD_KEY" \
   --secret-input-mode plaintext \
   --custom-compatibility openai \
   --gateway-port 18788 \

sandboxes/nemoclaw/nemoclaw-ui-extension/extension/index.ts

@@ -14,7 +14,7 @@ import "./styles.css";
 import { injectButton } from "./deploy-modal.ts";
 import { injectNavGroup, activateNemoPage, watchOpenClawNavClicks } from "./nav-group.ts";
 import { injectModelSelector, watchChatCompose } from "./model-selector.ts";
-import { ingestKeysFromUrl, DEFAULT_MODEL, resolveApiKey } from "./model-registry.ts";
+import { ingestKeysFromUrl, DEFAULT_MODEL, resolveApiKey, isKeyConfigured } from "./model-registry.ts";
 import { waitForClient, patchConfig, waitForReconnect } from "./gateway-bridge.ts";
 
 function inject(): boolean {
@@ -38,6 +38,22 @@ function watchGotoLinks() {
   });
 }
 
+/**
+ * Update the NemoClaw provider credential on the host so the sandbox
+ * proxy / inference router uses the real key for inference.local requests.
+ * Mirrors the policy-sync pattern in policy-page.ts.
+ */
+function injectKeyViaHost(key: string): void {
+  fetch("/api/inject-key", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ key }),
+  })
+    .then((r) => r.json())
+    .then((b) => console.log("[NeMoClaw] inject-key:", b))
+    .catch((e) => console.warn("[NeMoClaw] inject-key failed:", e));
+}
+
 /**
  * When API keys arrive via URL parameters (from the welcome UI), apply
  * the default model's provider config so the gateway has a valid key
@@ -103,8 +119,10 @@ function bootstrap() {
   watchChatCompose();
   watchGotoLinks();
 
-  if (keysIngested) {
+  const defaultKey = resolveApiKey(DEFAULT_MODEL.keyType);
+  if (keysIngested || isKeyConfigured(defaultKey)) {
     applyIngestedKeys();
+    injectKeyViaHost(defaultKey);
   }
 
   if (inject()) {

sandboxes/nemoclaw/nemoclaw-ui-extension/extension/model-registry.ts

@@ -45,8 +45,13 @@ export function setIntegrateApiKey(key: string): void {
   else localStorage.removeItem(LS_INTEGRATE_KEY);
 }
 
+const PLACEHOLDER_KEYS = ["not-used", "unused", "placeholder", "none", "null", "undefined"];
+
 export function isKeyConfigured(key: string): boolean {
-  return !!key && !key.startsWith("__");
+  if (!key || !key.trim()) return false;
+  const lower = key.trim().toLowerCase();
+  if (lower.startsWith("__")) return false;
+  return !PLACEHOLDER_KEYS.includes(lower);
 }
 
 /**