openzoo 0.48.2 → 0.48.4

package/lib/grokui.mjs

@@ -301,6 +301,35 @@ function newGroupThread(names) {
 const BIND_CHUNK_BYTES = 512 * 1024;
 // Chained auto-run commands per user message. Each hop is a paid call.
 const AUTO_MAX_STEPS = Number(process.env.OZ_AUTO_MAX_STEPS || 8);
+
+// Injected fresh on every AUTO turn, never persisted into the thread.
+//
+// The auto loop only continues while directives keep parsing, so a reply that
+// merely OFFERS ends the run — auto silently degrades to ask the moment the
+// model hedges. Observed live: "If you want, I can rewrite the prompt with
+// these fixes folded in", "Spawned mcp-integration — working on it" with
+// nothing spawned, and a user reduced to answering "no, this... impl all".
+// Models are trained to close on a consent question; in auto that instinct is
+// the bug. The system prompt is frozen into a thread at creation, so an
+// existing thread can only be reached by a per-turn message.
+const AUTO_DIRECTIVE = `AUTO MODE IS ON for this thread.
+
+Do the work in this turn. Do not ask whether to proceed, do not offer to do it,
+do not say what you are "about to" do and stop. The user has already consented
+by enabling auto — a question back to them is a dropped turn, and they must
+type "yes" to get what they already asked for.
+
+Concretely, NEVER end a turn with any of: "If you want, I can…", "Should I…?",
+"Let me know and I'll…", "Ready to proceed?", or a plan with no directive after
+it. If you catch yourself writing one, emit the RUN/WRITE/READ/SERVE/FETCH line
+instead — that IS the answer.
+
+Announcing an action does not perform it. "Spawned X", "working on it" and
+"kicked that off" are false unless the directive line is in this same reply.
+If a task needs several commands, emit the FIRST one now; you get its real
+output back and continue from there. Only stop to ask when the next step is
+genuinely destructive and irreversible, or when you truly cannot proceed
+without a fact only the user has.`;
 async function bindThread(t) {
   // Only bind what's NEW since the last successful bind, continuing the
   // existing context_id — previously this rebuilt and re-sent the WHOLE
@@ -378,8 +407,16 @@ 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);
@@ -391,9 +428,21 @@ 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)');
@@ -555,10 +604,16 @@ async function runTurn(threadId, userText, onEvent, images) {
   t.status = 'thinking';
   let reply = '';
   onEvent?.({ type: 'start', name: t.name, color: t.color });
+  // Transient: the nudge is appended for THIS call only and never pushed into
+  // t.messages, so it can't accumulate across a chained auto run or get bound
+  // into the thread's context.
+  const callMsgs = t.runMode === 'auto'
+    ? [...t.messages, { role: 'system', content: AUTO_DIRECTIVE }]
+    : t.messages;
   try {
     reply = onEvent
-      ? (await brainStream(t.messages, (delta) => onEvent({ type: 'delta', name: t.name, color: t.color, delta }), t.contextId)).trim()
-      : (await brain(t.messages, t.contextId)).trim();
+      ? (await brainStream(callMsgs, (delta) => onEvent({ type: 'delta', name: t.name, color: t.color, delta }), t.contextId)).trim()
+      : (await brain(callMsgs, t.contextId)).trim();
   } catch (e) {
     reply = `error: ${e.message}`;
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.2",
+  "version": "0.48.4",
   "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",