openzoo 0.48.1 → 0.48.3

package/lib/boxes.js

@@ -36,8 +36,15 @@ const CPU_FLAVORS = ['cpu3c', 'cpu3g', 'cpu5c'];
 // on musl they fail to build because the entrypoint has no toolchain, which is
 // why alpine boxes sat at 404 forever. Debian-slim runs the shim as published,
 // and its apt has a real sshd.
-const BOX_IMAGE = process.env.OPENZOO_BOX_IMAGE || 'node:22-slim';
-const BOX_PORTS = ['8402/http', '1337/http', '6080/http', '1340/http', '6081/http', '22/tcp'];
+// Everything the box runs is ALREADY IN THE IMAGE: the shim, its node_modules,
+// grokui, and sshd. Boxes used to apt-get and `npx openzoo` at boot, which put
+// a package fetch on the critical path from RunPod egress IPs — and those get
+// rate limited. A 429 from raw.githubusercontent arrives as a FILE whose
+// contents are the text "429: Too Many Requests", so node ran it and failed in
+// a way that read as an app bug rather than a download bug.
+// Still glibc, never musl: @solana's native deps don't build on the box.
+const BOX_IMAGE = process.env.OPENZOO_BOX_IMAGE || 'jrsdunn123/openzoo-box:latest';
+const BOX_PORTS = ['8402/http', '4173/http', '1337/http', '6080/http', '1340/http', '6081/http', '22/tcp'];
 const TTL_MAX_HOURS = 90 * 24;
 
 /**
@@ -50,18 +57,27 @@ const TTL_MAX_HOURS = 90 * 24;
  * host keys, and run sshd in the foreground of the background. Passwordless
  * root login is NOT enabled — key-only.
  *
- * `npx --yes openzoo` is still fetched at boot so a box always runs the
- * published shim; readyBox() polls /v1/models rather than assuming a timing.
+ * NOTHING IS FETCHED AT BOOT. sshd, the shim and grokui are baked into
+ * BOX_IMAGE by box.Dockerfile; this only copies /opt into the volume RunPod
+ * mounts over /workspace and starts things. readyBox() still polls /v1/models
+ * rather than assuming a timing.
  */
 const ENTRYPOINT = [
   'bash', '-lc',
   'set +e; mkdir -p /workspace /root/.openzoo /var/log/openzoo /run/sshd /root/.ssh; '
   + 'if [ -n "$OPENZOO_WALLET_JSON" ]; then printf %s "$OPENZOO_WALLET_JSON" > /root/.openzoo/wallet.json; chmod 600 /root/.openzoo/wallet.json; fi; '
-  + 'export DEBIAN_FRONTEND=noninteractive; apt-get update >/dev/null 2>&1; apt-get install -y --no-install-recommends openssh-server git ca-certificates >/dev/null 2>&1; '
   // RunPod hands the operator pubkey in one of these — accept whichever is set
   + 'for K in "$PUBLIC_KEY" "$SSH_PUBLIC_KEY"; do [ -n "$K" ] && echo "$K" >> /root/.ssh/authorized_keys; done; chmod 600 /root/.ssh/authorized_keys 2>/dev/null; '
   + 'ssh-keygen -A >/dev/null 2>&1; sed -i "s/^#\\?PermitRootLogin.*/PermitRootLogin prohibit-password/" /etc/ssh/sshd_config 2>/dev/null; /usr/sbin/sshd -e >/var/log/openzoo/sshd.log 2>&1 & '
-  + 'npx --yes openzoo > /var/log/openzoo/proxy.log 2>&1 & '
+  // RUN FROM /opt, never from /workspace. Staging the app onto the volume and
+  // gating on `[ ! -f .../bin/openzoo.js ]` is a trap: one interrupted copy
+  // leaves that file behind, every later boot skips the copy, and the box
+  // crash-loops on a half-copied node_modules forever (seen in production as
+  // "Cannot find module .../viem/accounts" every 15s, which also cost the pod
+  // its HTTP port mappings). The app is read-only at runtime; mutable state
+  // lives in /root/.openzoo.
+  + 'until OPENZOO_BIND=0.0.0.0 node /opt/openzoo/bin/openzoo.js >> /var/log/openzoo/proxy.log 2>&1; do echo "proxy exited, restarting" >> /var/log/openzoo/proxy.log; sleep 2; done & '
+  + 'until OZ_GROKUI_BIND=0.0.0.0 OZ_GROKUI_PORT=4173 node /opt/grokui/grokui.mjs >> /var/log/openzoo/grokui.log 2>&1; do echo "grokui exited, restarting" >> /var/log/openzoo/grokui.log; sleep 2; done & '
   // the capture agent answers the ports Grok Bot expects a Cursor sandbox on,
   // logging the protocol we do not yet speak (see lib/podagent.mjs)
   + 'if [ -n "$OZ_PODAGENT_B64" ]; then printf %s "$OZ_PODAGENT_B64" | base64 -d > /opt/podagent.mjs; node /opt/podagent.mjs > /var/log/openzoo/agent.log 2>&1 & fi; '

package/lib/grokui.mjs

@@ -14,6 +14,12 @@ import path from 'node:path';
 import { brain, brainStream, PROXY } from './podagent.mjs';
 
 const PORT = Number(process.env.OZ_GROKUI_PORT || 4173);
+// BIND HOST. Default 127.0.0.1 so the desktop app never exposes a shell-capable
+// UI to the LAN. A container MUST override it: binding the container's loopback
+// makes `docker run -p 4173:4173` refuse the connection, because the published
+// port forwards to the container's external interface, which nothing is on.
+// Mirrors OPENZOO_BIND in lib/proxy.js — same reasoning, same default.
+const BIND = process.env.OZ_GROKUI_BIND || process.env.OPENZOO_BIND || '127.0.0.1';
 const STORE_DIR = path.join(homedir(), '.openzoo');
 const STORE_FILE = path.join(STORE_DIR, 'grokui-threads.json');
 
@@ -372,8 +378,16 @@ function parseRun(reply) {
   // use U+FF5C FULLWIDTH VERTICAL LINE (｜) — matching only `|` parsed the
   // pretty-printed form in tests while missing what the model actually emits,
   // which is exactly how this survived one round of "fixed".
+  // The PARAMETER NAME is not fixed either. Models emit name="command",
+  // name="cmd", name="shell_command" and name="script" for the same thing —
+  // MEASURED live emitting `name="cmd"` inside an invoke named exec_command,
+  // against a parser that demanded name="command". One attribute apart, and
+  // the whole envelope was dropped in silence: the bot then explained what it
+  // was "about to run" forever, never running anything. Match the shape of the
+  // envelope, not one vendor's spelling of it.
   const SEP = '[|｜\\s]*';
-  const dsml = new RegExp(`<${SEP}DSML[^>]*\\bparameter\\b[^>]*name="command"[^>]*>([\\s\\S]*?)<\\/${SEP}DSML`, 'i').exec(reply);
+  const NAME = '(?:command|cmd|shell_command|script)';
+  const dsml = new RegExp(`<${SEP}DSML[^>]*\\bparameter\\b[^>]*\\bname="${NAME}"[^>]*>([\\s\\S]*?)<\\/${SEP}DSML`, 'i').exec(reply);
   if (dsml) return sanitizeRunCommand(dsml[1]);
 
   const m = /^[ \t>*-]*RUN:[ \t]*([\s\S]+)/m.exec(reply);
@@ -385,9 +399,21 @@ function parseRun(reply) {
   return sanitizeRunCommand(cmd);
 }
 
+// RUN through BASH, not /bin/sh. node's exec() defaults to /bin/sh, which on
+// Debian is dash — so every bash-ism a model writes (`for … do`, `[[ ]]`,
+// arrays, process substitution) dies as
+//   /bin/sh: 40: Syntax error: "do" unexpected
+// which reads as the model writing bad code when it wrote perfectly good bash.
+// Models overwhelmingly emit bash; give them bash. Windows is left alone so
+// node picks cmd.exe, and a box without /bin/bash falls back to the default.
+const RUN_SHELL = process.platform !== 'win32' && existsSync('/bin/bash') ? '/bin/bash' : undefined;
+// Installing a toolchain (apt-get, pip, cargo) routinely outruns two minutes,
+// and a killed install leaves a half-configured box that fails confusingly.
+const RUN_TIMEOUT_MS = Number(process.env.OZ_RUN_TIMEOUT_MS || 600000);
+
 function execCommand(command, cwd) {
   return new Promise((resolve) => {
-    exec(command, { cwd, timeout: 120000, maxBuffer: 10 * 1024 * 1024 }, (err, stdout, stderr) => {
+    exec(command, { cwd, shell: RUN_SHELL, timeout: RUN_TIMEOUT_MS, maxBuffer: 10 * 1024 * 1024 }, (err, stdout, stderr) => {
       let out = (stdout || '') + (stderr ? '\n' + stderr : '');
       if (err) out += `\n(exit ${err.code ?? 1})`;
       resolve(out.slice(0, 6000) || '(no output)');
@@ -1519,4 +1545,4 @@ const server = http.createServer((req, res) => {
   res.end(APP_HTML);
 });
 
-server.listen(PORT, '127.0.0.1', () => console.log(`[grokui] http://localhost:${PORT}`));
+server.listen(PORT, BIND, () => console.log(`[grokui] http://${BIND === '0.0.0.0' ? 'localhost' : BIND}:${PORT}`));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.1",
+  "version": "0.48.3",
   "description": "Local x402-paying proxy + MCP server for openzoo.fun — point any OpenAI-compatible harness (Cursor, Claude Code, aider, SDKs) at localhost and it pays per call from a local burner wallet. Solana and Base rails live; Robinhood experimental.",
   "license": "MIT",
   "type": "module",