fix: add local-inference policy preset for Ollama/vLLM host access (#693)

bin/lib/onboard.js

@@ -3221,11 +3221,19 @@ async function setupOpenclaw(sandboxName, model, provider) {
 // ── Step 7: Policy presets ───────────────────────────────────────
 
 // eslint-disable-next-line complexity
-async function _setupPolicies(sandboxName) {
+async function _setupPolicies(sandboxName, provider = null) {
   step(7, 7, "Policy presets");
 
   const suggestions = ["pypi", "npm"];
 
+  // Auto-detect local inference — sandbox needs host gateway egress
+  const sandbox = registry.getSandbox(sandboxName);
+  const sandboxProvider = provider || (sandbox ? sandbox.provider : null);
+  if (sandboxProvider === "ollama-local" || sandboxProvider === "vllm-local") {
+    suggestions.push("local-inference");
+    console.log(`  Auto-detected: ${sandboxProvider} → suggesting local-inference preset`);
+  }
+
   // Auto-detect based on env tokens
   if (getCredential("TELEGRAM_BOT_TOKEN")) {
     suggestions.push("telegram");

bin/lib/runner.js

@@ -8,6 +8,33 @@ const { detectDockerHost } = require("./platform");
 const ROOT = path.resolve(__dirname, "..", "..");
 const SCRIPTS = path.join(ROOT, "scripts");
 
+/**
+ * Redact known secret patterns from a string to prevent credential leaks
+ * in logs, error messages, and terminal output.
+ *
+ * Matches:
+ *  - Environment-style assignments: NVIDIA_API_KEY=sk-... → NVIDIA_API_KEY=<REDACTED>
+ *  - NVIDIA API key prefix:         nvapi-Abc123...       → <REDACTED>
+ *  - GitHub PAT prefix:             ghp_Abc123...         → <REDACTED>
+ *  - Bearer tokens:                 Bearer eyJhb...       → Bearer <REDACTED>
+ */
+const SECRET_PATTERNS = [
+  { re: /(NVIDIA_API_KEY|API_KEY|TOKEN|SECRET|PASSWORD|CREDENTIAL|_KEY)=\S+/gi, replacement: "$1=<REDACTED>" },
+  { re: /nvapi-[A-Za-z0-9_-]{10,}/g, replacement: "<REDACTED>" },
+  { re: /ghp_[A-Za-z0-9]{30,}/g, replacement: "<REDACTED>" },
+  { re: /(Bearer )\S+/gi, replacement: "$1<REDACTED>" },
+];
+
+function redactSecrets(str) {
+  if (typeof str !== "string") return str;
+  let result = str;
+  for (const { re, replacement } of SECRET_PATTERNS) {
+    re.lastIndex = 0;
+    result = result.replace(re, replacement);
+  }
+  return result;
+}
+
 const dockerHost = detectDockerHost();
 if (dockerHost) {
   process.env.DOCKER_HOST = dockerHost.dockerHost;
@@ -22,7 +49,7 @@ function run(cmd, opts = {}) {
     env: { ...process.env, ...opts.env },
   });
   if (result.status !== 0 && !opts.ignoreError) {
-    console.error(`  Command failed (exit ${result.status}): ${cmd.slice(0, 80)}`);
+    console.error(`  Command failed (exit ${result.status}): ${redactSecrets(cmd.slice(0, 80))}`);
     process.exit(result.status || 1);
   }
   return result;
@@ -37,7 +64,7 @@ function runInteractive(cmd, opts = {}) {
     env: { ...process.env, ...opts.env },
   });
   if (result.status !== 0 && !opts.ignoreError) {
-    console.error(`  Command failed (exit ${result.status}): ${cmd.slice(0, 80)}`);
+    console.error(`  Command failed (exit ${result.status}): ${redactSecrets(cmd.slice(0, 80))}`);
     process.exit(result.status || 1);
   }
   return result;
@@ -85,4 +112,4 @@ function validateName(name, label = "name") {
   return name;
 }
 
-module.exports = { ROOT, SCRIPTS, run, runCapture, runInteractive, shellQuote, validateName };
+module.exports = { ROOT, SCRIPTS, run, runCapture, runInteractive, shellQuote, validateName, redactSecrets };

bin/nemoclaw.js

@@ -632,8 +632,10 @@ async function setup() {
 }
 
 async function setupSpark() {
-  // setup-spark.sh configures Docker cgroups — it does not use NVIDIA_API_KEY.
-  run(`sudo bash "${SCRIPTS}/setup-spark.sh"`);
+  await ensureApiKey();
+  run(`sudo -E bash "${SCRIPTS}/setup-spark.sh"`, {
+    env: { NVIDIA_API_KEY: process.env.NVIDIA_API_KEY },
+  });
 }
 
 // eslint-disable-next-line complexity

install.sh

@@ -679,7 +679,14 @@ install_nemoclaw() {
       || warn "Pre-extraction failed — npm install may fail if openclaw tarball is broken"
     spin "Installing NemoClaw dependencies" npm install --ignore-scripts
     spin "Building NemoClaw plugin" bash -c 'cd nemoclaw && npm install --ignore-scripts && npm run build'
-    spin "Linking NemoClaw CLI" npm link
+    # Use sudo for npm link only when the global prefix is not writable
+    local npm_global_prefix
+    npm_global_prefix="$(npm config get prefix 2>/dev/null)" || true
+    local sudo_cmd=""
+    if [ -n "$npm_global_prefix" ] && [ ! -w "$npm_global_prefix" ] && [ "$(id -u)" -ne 0 ]; then
+      sudo_cmd="sudo"
+    fi
+    spin "Linking NemoClaw CLI" $sudo_cmd npm link
   else
     info "Installing NemoClaw from GitHub…"
     # Resolve the latest release tag so we never install raw main.
@@ -697,7 +704,14 @@ install_nemoclaw() {
       || warn "Pre-extraction failed — npm install may fail if openclaw tarball is broken"
     spin "Installing NemoClaw dependencies" bash -c "cd \"$nemoclaw_src\" && npm install --ignore-scripts"
     spin "Building NemoClaw plugin" bash -c "cd \"$nemoclaw_src\"/nemoclaw && npm install --ignore-scripts && npm run build"
-    spin "Linking NemoClaw CLI" bash -c "cd \"$nemoclaw_src\" && npm link"
+    # Use sudo for npm link only when the global prefix is not writable
+    local npm_global_prefix
+    npm_global_prefix="$(npm config get prefix 2>/dev/null)" || true
+    local sudo_cmd=""
+    if [ -n "$npm_global_prefix" ] && [ ! -w "$npm_global_prefix" ] && [ "$(id -u)" -ne 0 ]; then
+      sudo_cmd="sudo"
+    fi
+    spin "Linking NemoClaw CLI" bash -c "cd \"$nemoclaw_src\" && $sudo_cmd npm link"
   fi
 
   refresh_path

nemoclaw-blueprint/policies/presets/local-inference.yaml

@@ -0,0 +1,28 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+preset:
+  name: local-inference
+  description: "Local inference access (Ollama, vLLM) via host gateway"
+
+network_policies:
+  local_inference:
+    name: local_inference
+    endpoints:
+      - host: host.openshell.internal
+        port: 11434
+        protocol: rest
+        enforcement: enforce
+        rules:
+          - allow: { method: GET, path: "/**" }
+          - allow: { method: POST, path: "/**" }
+      - host: host.openshell.internal
+        port: 8000
+        protocol: rest
+        enforcement: enforce
+        rules:
+          - allow: { method: GET, path: "/**" }
+          - allow: { method: POST, path: "/**" }
+    binaries:
+      - { path: /usr/local/bin/openclaw }
+      - { path: /usr/local/bin/claude }

scripts/telegram-bridge.js

@@ -105,13 +105,13 @@ function runAgentInSandbox(message, sessionId) {
     const confPath = `${confDir}/config`;
     require("fs").writeFileSync(confPath, sshConfig, { mode: 0o600 });
 
-    // Pass message and API key via stdin to avoid shell interpolation.
-    // The remote command reads them from environment/stdin rather than
-    // embedding user content in a shell string.
+    // Pass API key via SendEnv + ssh config to avoid exposing it in
+    // process arguments (visible in ps aux / /proc/*/cmdline).
     const safeSessionId = String(sessionId).replace(/[^a-zA-Z0-9-]/g, "");
-    const cmd = `export NVIDIA_API_KEY=${shellQuote(API_KEY)} && nemoclaw-start openclaw agent --agent main --local -m ${shellQuote(message)} --session-id ${shellQuote("tg-" + safeSessionId)}`;
+    const cmd = `nemoclaw-start openclaw agent --agent main --local -m ${shellQuote(message)} --session-id ${shellQuote("tg-" + safeSessionId)}`;
 
-    const proc = spawn("ssh", ["-T", "-F", confPath, `openshell-${SANDBOX}`, cmd], {
+    const proc = spawn("ssh", ["-T", "-F", confPath, "-o", "SendEnv=NVIDIA_API_KEY", `openshell-${SANDBOX}`, cmd], {
+      env: { ...process.env, NVIDIA_API_KEY: API_KEY },
       timeout: 120000,
       stdio: ["ignore", "pipe", "pipe"],
     });

scripts/walkthrough.sh

@@ -84,10 +84,9 @@ tmux kill-session -t "$SESSION" 2>/dev/null || true
 # Create session with TUI on the left
 tmux new-session -d -s "$SESSION" -x 200 -y 50 "openshell term"
 
-# Split right pane for the agent
-# NVIDIA_API_KEY is not needed inside the sandbox — inference is proxied
-# through the OpenShell gateway which injects credentials server-side.
-tmux split-window -h -t "$SESSION" \
+# Split right pane for the agent — pass API key via env to avoid leaking
+# the secret in process arguments (visible in ps aux / /proc/*/cmdline).
+tmux split-window -h -t "$SESSION" -e "NVIDIA_API_KEY=$NVIDIA_API_KEY" \
   "openshell sandbox connect nemoclaw -- bash -c 'nemoclaw-start openclaw agent --agent main --local --session-id live'"
 
 # Even split

spark-install.md

@@ -208,7 +208,18 @@ The OpenClaw gateway includes a built-in web UI. Access it at:
 http://127.0.0.1:18789/#token=<your-gateway-token>
 ```
 
-Find your gateway token in `~/.openclaw/openclaw.json` under `gateway.auth.token` inside the sandbox.
+To retrieve your gateway token, run this from inside the sandbox:
+
+```console
+$ jq -r '.gateway.auth.token' /sandbox/.openclaw/openclaw.json
+```
+
+Or from the host, download the config first:
+
+```console
+$ openshell sandbox download <sandbox-name> /sandbox/.openclaw/openclaw.json ./openclaw.json
+$ jq -r '.gateway.auth.token' ./openclaw.json
+```
 
 > **Important**: Use `127.0.0.1` (not `localhost`) — the gateway's origin check requires an exact match. External dashboards like Mission Control cannot currently connect due to the gateway resetting `controlUi.allowedOrigins` on every config reload (see [openclaw#49950](https://github.com/openclaw/openclaw/issues/49950)).
 

test/policies.test.js

@@ -5,11 +5,13 @@ import assert from "node:assert/strict";
 import { describe, it, expect } from "vitest";
 import policies from "../bin/lib/policies";
 
+const expectedPresets = ["discord", "docker", "huggingface", "jira", "local-inference", "npm", "outlook", "pypi", "slack", "telegram"];
+
 describe("policies", () => {
   describe("listPresets", () => {
-    it("returns all 9 presets", () => {
+    it(`returns all ${expectedPresets.length} presets`, () => {
       const presets = policies.listPresets();
-      expect(presets.length).toBe(9);
+      expect(presets.length).toBe(expectedPresets.length);
     });
 
     it("each preset has name and description", () => {
@@ -20,22 +22,8 @@ describe("policies", () => {
     });
 
     it("returns expected preset names", () => {
-      const names = policies
-        .listPresets()
-        .map((p) => p.name)
-        .sort();
-      const expected = [
-        "discord",
-        "docker",
-        "huggingface",
-        "jira",
-        "npm",
-        "outlook",
-        "pypi",
-        "slack",
-        "telegram",
-      ];
-      expect(names).toEqual(expected);
+      const names = policies.listPresets().map((p) => p.name).sort();
+      expect(names).toEqual(expectedPresets);
     });
   });
 
@@ -296,6 +284,33 @@ describe("policies", () => {
     });
   });
 
+  describe("local-inference preset", () => {
+    it("loads and contains host.openshell.internal", () => {
+      const content = policies.loadPreset("local-inference");
+      expect(content).toBeTruthy();
+      const hosts = policies.getPresetEndpoints(content);
+      expect(hosts.includes("host.openshell.internal")).toBeTruthy();
+    });
+
+    it("allows Ollama port 11434 and vLLM port 8000", () => {
+      const content = policies.loadPreset("local-inference");
+      expect(content.includes("port: 11434")).toBe(true);
+      expect(content.includes("port: 8000")).toBe(true);
+    });
+
+    it("has a binaries section", () => {
+      const content = policies.loadPreset("local-inference");
+      expect(content.includes("binaries:")).toBe(true);
+    });
+
+    it("extracts valid network_policies entries", () => {
+      const content = policies.loadPreset("local-inference");
+      const entries = policies.extractPresetEntries(content);
+      expect(entries).toBeTruthy();
+      expect(entries.includes("local_inference")).toBe(true);
+    });
+  });
+
   describe("preset YAML schema", () => {
     it("no preset has rules at NetworkPolicyRuleDef level", () => {
       // rules must be inside endpoints, not as sibling of endpoints/binaries