@tiens.nguyen/gonext-local-worker 1.0.259 → 1.0.261

package/gonext-repl.mjs

@@ -1025,13 +1025,20 @@ const normalizeThought = (buf) => {
   const t = String(buf || "");
   const ci = t.search(/<code[\s>]/i);
   const head = ci >= 0 ? t.slice(0, ci) : t;
-  // 1) a genuine Thought prose line (before the code) — only if it isn't itself code.
+  // 1) a genuine Thought prose line (before the code) — only if it isn't itself code AND
+  //    reads like an actual clause. A streaming token boundary can leave the first
+  //    non-empty line as a single bare word (e.g. "string", when the model is rambling
+  //    about "multi-line strings"); shown alone on the blinking line it's meaningless and,
+  //    because liveThought only updates on a truthy value, it can FREEZE there through a
+  //    long silent prompt-eval ("string… (777s)"). Require ≥2 words and a little length so
+  //    a lone fragment falls through to the tool label / "Thinking…" instead.
   const prose = head
     .replace(/^\s*Thought:\s*/i, "")
     .split(/\n/)
     .map((x) => x.trim())
     .find(Boolean);
-  if (prose && !_thoughtLooksCodey(prose)) return prose.slice(0, 100);
+  const proseIsClause = !!prose && prose.length >= 6 && prose.trim().split(/\s+/).length >= 2;
+  if (proseIsClause && !_thoughtLooksCodey(prose)) return prose.slice(0, 100);
   // 2) otherwise name the tool being called (from the code part).
   const codePart = ci >= 0 ? t.slice(ci) : t;
   const m = /(?:<code>\s*)?\b([a-z_]\w*)\s*\(\s*(?:path\s*=\s*)?["']?([^"'\n,)]*)/i.exec(codePart);
@@ -1117,6 +1124,7 @@ async function runAgentTurn(history) {
   // "◐ working… (7s)" — instead of dumping the model's rambling reasoning and stamping a
   // "✔ completed thought" on every token-stream micro-pause (the reported clutter).
   const QUIET_MS = 700; // stream idle this long ⇒ we're waiting on the model
+  const STALE_THOUGHT_MS = 25000; // no new tokens this long ⇒ the current Thought clause is stale (#95 A)
   let lastContentAt = Date.now();
   let thinking = false;
   let thinkingSince = 0;
@@ -1154,6 +1162,11 @@ async function runAgentTurn(history) {
       thinking = true; thinkingSince = lastContentAt; thinkWord = pickWord(); lastWordBucket = 0;
     }
     const now = Date.now();
+    // Age out a stale Thought clause: if the model has produced NO new stream content for a
+    // while (e.g. a 150s+ silent prompt-eval on a huge context), the last clause is no longer
+    // "what it's doing now" — drop it so the line falls back to the playful word / "Thinking…"
+    // instead of freezing on an old thought while the timer climbs ("string… (777s)", #95 A).
+    if (liveThought && now - lastContentAt > STALE_THOUGHT_MS) liveThought = "";
     const secs = thinkSecs();
     // Rotate the playful word once per 12s window (guarded so the fast ticker doesn't
     // re-roll it many times within the same second).
@@ -1750,6 +1763,18 @@ async function main() {
         console.log(shown ? "" : `\n\n${answer}\n`);
       } else {
         history.pop(); // aborted/failed/cancelled turn — don't poison the next request
+        // A turn that ended WITHOUT an answer (Ctrl+C cancel, or an abort/failure) must not
+        // silently launch another turn. Any keystrokes the user typed DURING the dead turn
+        // sit in `_lineQueue` — they land there (not ignored) whenever `following` was still
+        // false, e.g. during the agent-ask round-trip at the very start of the turn, or
+        // while the user was mashing "continue"/Enter. If left queued, `nextLine()` drains
+        // them back-to-back on the next loop iterations, each auto-firing a fresh turn that
+        // a straggler Ctrl+C then lands on — the reported "every turn immediately shows
+        // (cancelled)" loop. Drop that buffered input so the NEXT turn only starts when the
+        // user deliberately types something new (bug #92). Resume via "continue" still works:
+        // that's typed AFTER the cancel note, so it arrives through the live waiter, not the
+        // queue.
+        if (_lineQueue.length > 0) _lineQueue.length = 0;
         // Ctrl+C cancel already printed its own "(cancelled)" note in runAgentTurn — a
         // second "(no answer)" line here would be redundant/misleading.
         if (cancelled) {
@@ -1763,6 +1788,9 @@ async function main() {
       }
     } catch (err) {
       history.pop();
+      // Same reasoning as the no-answer branch (bug #92): don't let input buffered during a
+      // turn that just threw auto-fire the next one.
+      if (_lineQueue.length > 0) _lineQueue.length = 0;
       console.error(red(`\ngonext: ${err.message}\n`));
     }
   }

package/gonext_agent_chat.py

@@ -1077,6 +1077,35 @@ def _plain_reply(messages: list, base_url: str, api_key: str, model_id: str,
     raise RuntimeError(f"couldn't get a response ({_clip(str(last_err), 160)})")
 
 
+# A model-server call can fail for two very different reasons, and they deserve
+# different handling:
+#   • the BACKEND is unavailable — an nginx 502/503/504 (the upstream Ollama/MLX box is
+#     down, restarting, still loading the weights, or overloaded past the proxy's read
+#     timeout), or the socket refused/timed out. These are INFRA outages: retrying a bit
+#     harder can ride out a blip, and if it persists the honest thing is to SAY the
+#     coding backend is down — never to hallucinate a generic answer as if we succeeded.
+#   • a genuine model/agent error — worth the plain-reply degrade.
+# `str(e)` for a proxy error is the raw nginx HTML ("<html>…502 Bad Gateway…"), so match
+# on the status text/codes and the common connection-failure phrases.
+def _is_backend_unavailable(err) -> bool:
+    s = str(err or "").lower()
+    return (
+        "502 bad gateway" in s
+        or "503 service" in s
+        or "504 gateway" in s
+        or "bad gateway" in s
+        or "gateway time-out" in s
+        or "gateway timeout" in s
+        or "connection refused" in s
+        or "connection reset" in s
+        or "connection aborted" in s
+        or "failed to establish a new connection" in s
+        or "max retries exceeded" in s
+        or "read timed out" in s
+        or "timed out" in s
+    )
+
+
 def _synthesize_document(gathered: list, task: str, base_url: str, api_key: str,
                          model_id: str) -> str:
     """Turn raw gathered research observations into ONE clean, organized document body
@@ -1179,11 +1208,19 @@ def _normalize_code_tags(text: str) -> str:
     return fixed
 
 
-def _escape_raw_newlines_in_strings(code: str) -> str:
+def _escape_raw_newlines_in_strings(code: str, defuse_tags: bool = False) -> str:
     """Turn RAW newlines/tabs that sit INSIDE a single/double-quoted string literal into
     \\n/\\t escapes. Triple-quoted strings, comments, and everything outside a string are
     copied verbatim. A tiny hand-rolled scanner (the code doesn't parse, so ast/tokenize
-    can't help). Used only by _repair_unterminated_string below."""
+    can't help). Used by _repair_unterminated_string below.
+
+    defuse_tags: also rewrite a literal <code>/</code> that appears INSIDE a string literal
+    (e.g. the file content the model is writing legitimately contains React's default
+    "Edit <code>src/App.js</code>") to <co\\x64e>/</co\\x64e>. The `\\x64` is Python's hex
+    escape for 'd', so at RUNTIME the string is still exactly "<code>"/"</code>" — but the
+    literal source no longer contains the token smolagents uses as its code-block delimiter,
+    so its parser can't mis-close the block on content. Only touches tags inside strings;
+    the real closing </code> lives OUTSIDE any string and is left intact."""
     out = []
     i, n = 0, len(code)
     quote = None      # the delimiter of the string we're inside, or None
@@ -1196,7 +1233,10 @@ def _escape_raw_newlines_in_strings(code: str) -> str:
                 end = code.find(three, i + 3)
                 if end == -1:
                     out.append(code[i:]); break
-                out.append(code[i:end + 3]); i = end + 3; continue
+                seg = code[i:end + 3]
+                if defuse_tags:
+                    seg = seg.replace("</code>", "</co\\x64e>").replace("<code>", "<co\\x64e>")
+                out.append(seg); i = end + 3; continue
             if ch == "#":                          # comment — copy to end of line
                 j = code.find("\n", i)
                 if j == -1:
@@ -1212,6 +1252,11 @@ def _escape_raw_newlines_in_strings(code: str) -> str:
                 out.append(ch); escaped = True; i += 1; continue
             if ch == quote:
                 out.append(ch); quote = None; i += 1; continue
+            # Defuse a delimiter-colliding tag INSIDE the string (see docstring).
+            if defuse_tags and code.startswith("</code>", i):
+                out.append("</co\\x64e>"); i += 7; continue
+            if defuse_tags and code.startswith("<code>", i):
+                out.append("<co\\x64e>"); i += 6; continue
             out.append({"\n": "\\n", "\r": "\\r", "\t": "\\t"}.get(ch, ch)); i += 1
     return "".join(out)
 
@@ -1242,15 +1287,42 @@ def _repair_unterminated_string(code: str) -> str:
 def _repair_code_block_strings(text: str) -> str:
     """Apply _repair_unterminated_string to the python inside a <code>…</code> block
     (leaving the surrounding Thought prose untouched). No-op when there's no block or the
-    block already parses."""
-    m = re.search(r"<code>([\s\S]*?)</code>", text)
-    if not m:
+    block already parses.
+
+    Two-stage rescue:
+      1. NON-GREEDY span (first <code> → first </code>) + raw-newline escaping — the common
+         case (multi-line file content with unescaped newlines).
+      2. If that still won't compile AND the content itself contains extra </code> tags
+         (the model is writing a file whose CONTENT legitimately has <code>/</code> — e.g.
+         React's default App.js "Edit <code>src/App.js</code>"), the non-greedy match closed
+         the block early on that inner tag → a truncated, unterminatable fragment. Retry with
+         the GREEDY span (first <code> → LAST </code>) and defuse the in-string tags so the
+         only real </code> left is the true closer. This is the fix for the create_file loop
+         where the coder burned minutes/100k+ tokens on 'unterminated string literal'."""
+    om = re.search(r"<code>", text)
+    if not om:
         return text
-    inner = m.group(1)
-    fixed = _repair_unterminated_string(inner)
-    if fixed == inner:
+    open_end = om.end()
+    cm = re.search(r"</code>", text[open_end:])
+    if not cm:
         return text
-    return text[:m.start(1)] + fixed + text[m.end(1):]
+    close_start = open_end + cm.start()
+    inner = text[open_end:close_start]
+    fixed = _repair_unterminated_string(inner)
+    if fixed != inner:
+        return text[:open_end] + fixed + text[close_start:]
+    # Stage 2 — only worth trying when there's MORE than one </code> (extra tags in content).
+    last = text.rfind("</code>")
+    if last <= close_start:
+        return text  # single </code>; nothing a greedy re-read would change
+    greedy = text[open_end:last]
+    sanitized = _escape_raw_newlines_in_strings(greedy, defuse_tags=True)
+    try:
+        compile(sanitized, "<gonext-code>", "exec")
+    except Exception:  # noqa: BLE001
+        return text  # still not parseable — leave it for the loop-breaker / model retry
+    # Rebuild so the FIRST literal </code> is now the true closer (inner ones are defused).
+    return text[:open_end] + sanitized + text[last:]
 
 
 def _strip_think(text: str) -> str:
@@ -4173,6 +4245,12 @@ def run_agent_chat(cfg):
             # Largest SINGLE-request prompt-token count seen this turn (task #84) — the
             # peak context size, not the sum. The REPL keeps a per-workspace max of this.
             self._code_tokens_max = 0
+            # Loop-breaker (#95 B1): how many CONSECUTIVE generations have produced a code
+            # block that STILL won't compile after our repair. A weak coder can get stuck
+            # re-emitting the same unparseable tool call (seen live: create_file with a
+            # multi-line string, ~10 identical retries, 863s/133k tokens). After a few in a
+            # row we stop feeding smolagents an un-runnable block and finish honestly.
+            self._parse_fail_streak = 0
             # stream_agent: request the completion with stream=True and reassemble the
             # deltas into one message. Enabled for Ollama, where a long single-shot
             # (non-stream) generation sends nothing until the very end — a reverse proxy
@@ -4445,9 +4523,16 @@ def run_agent_chat(cfg):
             hb.start()
             last_err = None
             try:
-                for attempt in range(3):
+                # A gateway outage (502/503/504, refused/reset/timeout) is transient infra —
+                # the remote box may just be cold or the proxy briefly down — so give it MORE
+                # attempts and a longer backoff to ride out a blip, instead of the 3 quick
+                # tries a normal error gets. The attempt count is decided per-error: we start
+                # with the generic budget and extend it the moment we see a gateway error.
+                attempt, max_attempts = 0, 3
+                while attempt < max_attempts:
                     try:
                         msg = self._generate_once(*args, **kwargs)
+                        last_err = None
                         break
                     except Exception as e:  # noqa: BLE001
                         emsg = str(e)
@@ -4465,10 +4550,19 @@ def run_agent_chat(cfg):
                                 "names (for Ollama use the tag, e.g. qwen3:14b)."
                             ) from e
                         last_err = e
-                        if attempt < 2:
-                            _log(f"generate transient error (attempt {attempt + 1}/3): {e} — retrying")
-                            time.sleep(1.5 * (attempt + 1))
-                else:
+                        gateway = _is_backend_unavailable(e)
+                        if gateway and max_attempts < 6:
+                            max_attempts = 6  # stretch the budget for a flaky backend
+                        attempt += 1
+                        if attempt < max_attempts:
+                            # Longer, capped backoff for gateway errors (up to ~8s) so a
+                            # cold/loading model has time to come back; short for the rest.
+                            delay = min(8.0, 2.0 * attempt) if gateway else 1.5 * attempt
+                            kind = "gateway/backend unavailable" if gateway else "transient"
+                            _log(f"generate {kind} error (attempt {attempt}/{max_attempts}): "
+                                 f"{_clip(emsg, 160)} — retrying in {delay:.0f}s")
+                            time.sleep(delay)
+                if last_err is not None:
                     raise last_err
             finally:
                 hb_stop.set()
@@ -4532,7 +4626,42 @@ def run_agent_chat(cfg):
                             _log("repaired unterminated string literal in tool-call code")
                         if repaired != content:
                             msg.content = repaired
+                        # Loop-breaker (#95 B1): if the block STILL won't compile after the
+                        # repair, the model is stuck emitting an unparseable tool call. Count
+                        # consecutive failures; after a few, finish honestly rather than burn
+                        # the whole step budget — each retry is a full (often 150s+) model
+                        # call on an ever-growing context. B2's greedy/defuse fix means the
+                        # <code>-in-content collision now compiles here, so this only trips on
+                        # genuinely un-fixable code (e.g. a truncated generation).
+                        _cb = re.search(r"<code>([\s\S]*?)</code>", msg.content or "")
+                        _still_bad = False
+                        if _cb:
+                            try:
+                                compile(_cb.group(1), "<gonext-code>", "exec")
+                            except Exception:  # noqa: BLE001
+                                _still_bad = True
+                        if _still_bad:
+                            self._parse_fail_streak += 1
+                            _log("tool-call code still unparseable after repair "
+                                 f"(streak {self._parse_fail_streak}/3)")
+                            if self._parse_fail_streak >= 3:
+                                _log("parse-fail streak hit limit → ending the turn honestly "
+                                     "instead of looping on an unparseable tool call")
+                                _stuck_msg = (
+                                    "I got stuck on a code-formatting error: my edit to the "
+                                    "file kept failing to parse (a string/escaping issue in the "
+                                    "tool call), so I stopped instead of retrying in a loop. "
+                                    "Any dev server I started is still running. Please try "
+                                    "again, or tell me the specific change to make and I'll "
+                                    "keep it to a small, single-line edit."
+                                )
+                                msg.content = ("<code>\nfinal_answer("
+                                               + repr(_stuck_msg) + ")\n</code>")
+                                self._parse_fail_streak = 0
+                        else:
+                            self._parse_fail_streak = 0
                     else:
+                        self._parse_fail_streak = 0  # a prose turn isn't a parse failure
                         stripped = content.strip()
                         looks_final = (
                             len(stripped) >= 400 or re.search(
@@ -5778,6 +5907,26 @@ def run_agent_chat(cfg):
             _log(f"agent config error (not degrading): {cfg_err}")
             _emit({"type": "final", "text": f"⚠️ {cfg_err}"})
             return
+        # Coding-model BACKEND outage (502/503/504, refused/timeout): the tool-capable
+        # loop never got to run because its code model was unreachable. Do NOT fall through
+        # to _plain_reply — a bare chat model would stream a confident GENERIC how-to (e.g.
+        # "to create a React app, run npx create-react-app…") that looks like a real answer
+        # but did NOTHING the user asked (seen live: "create a reactjs and start it" → the
+        # coder 502'd 3× → generic instructions, no warning). That's worse than an honest
+        # failure: it hides the outage and, being persisted to history, misleads the next
+        # turn. Emit a clear, standalone "backend is down, retry" and stop — the checkpoint
+        # is kept (not cleared) so a later "continue" resumes once the box is back.
+        if _is_backend_unavailable(e):
+            _log(f"agent error (coding-model backend unavailable, NOT degrading): {_clip(str(e), 160)}")
+            _emit({"type": "final", "text": (
+                "⚠️ The coding model backend is currently unavailable (the server returned "
+                "a gateway error / timed out), so I couldn't run the tools to do this. It's "
+                "usually starting up, reloading the model, or briefly overloaded — not a "
+                "problem with your request. Please try again in a moment; if it keeps "
+                "happening, check that the coding model server is up. (I didn't make any "
+                "changes.)"
+            )})
+            return
         # An unexpected failure inside the agent loop should still return a useful
         # answer from the conversation rather than surfacing a raw error to the user —
         # but only when the fallback ACTUALLY has something useful to say.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.259",
+  "version": "1.0.261",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",