npm - @tiens.nguyen/gonext-local-worker - Versions diffs - 1.0.326 → 1.0.328 - Mend

@tiens.nguyen/gonext-local-worker 1.0.326 → 1.0.328

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (3) hide show

package/gonext-repl.mjs CHANGED Viewed

@@ -456,6 +456,20 @@ function slashCloseSeq(line, cursor, cols) {
     `\r\x1b[${Math.min(width, PROMPT_COLS + cursor + 1)}G`
   );
 }
+// A window resize reflows every row, so the overlay's row arithmetic (and the width it
+// clipped to) is stale the moment it happens: rows that fit before may now wrap, which is
+// precisely what made the menu stack in the first place (#109 v3). Redraw at the new width
+// — the render ends with ESC[J, so a block that used to be taller leaves no orphans behind.
+if (process.stdout.isTTY) {
+  process.stdout.on("resize", () => {
+    if (!slashOverlayActive) return;
+    try {
+      renderSlashOverlay(slashSel);
+    } catch {
+      closeSlashOverlay(); // never let a resize wedge the prompt
+    }
+  });
+}
 if (process.stdin.isTTY) {
   process.stdin.on("keypress", (_ch, key) => {
     // An arrow-key list picker (/server) is up → ↑/↓ move the highlight, Enter chooses.

package/gonext_agent_chat.py CHANGED Viewed

@@ -1469,6 +1469,28 @@ def _plain_reply(messages: list, base_url: str, api_key: str, model_id: str,
 #   • 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 _http_status_of(err):
+    """The HTTP status an exception carries, or None if it isn't an HTTP error.
+    openai-python's APIStatusError exposes `.status_code` (and `.response.status_code`).
+    smolagents wraps model exceptions in AgentGenerationError, so the chain is walked —
+    otherwise every error reaching the turn-level handler would look status-less and fall
+    back to matching on message text."""
+    cur, depth = err, 0
+    while cur is not None and depth < 6:
+        for attr in ("status_code", "http_status", "status"):
+            code = getattr(cur, attr, None)
+            if isinstance(code, int) and 100 <= code <= 599:
+                return code
+        resp = getattr(cur, "response", None)
+        code = getattr(resp, "status_code", None)
+        if isinstance(code, int) and 100 <= code <= 599:
+            return code
+        cur = (getattr(cur, "__cause__", None) or getattr(cur, "__context__", None))
+        depth += 1
+    return None
 def _is_backend_overloaded(err) -> bool:
     """True when the model server ACCEPTED the connection but REFUSED the work because it
     is saturated or we are over a quota — 429 / engine_overloaded / rate limit (task #114).
@@ -1481,7 +1503,15 @@ def _is_backend_overloaded(err) -> bool:
     Live trigger: Moonshot returned
       Error code: 429 - {'error': {'message': 'The engine is currently overloaded, please
       try again later', 'type': 'engine_overloaded_error'}}
-    A local Ollama/MLX box does not produce these, so nothing keyed off this fires there."""
+    A local Ollama/MLX box does not produce these, so nothing keyed off this fires there.
+    A STRUCTURED status wins when the exception carries one: a real 429 is an overload
+    whatever its body says, and a real 401/500/... is NOT one even if some tool output
+    quoted the words "rate limit" into the message. Only when there is no status at all
+    (a bare string, a wrapped error that lost its cause) do we fall back to matching text."""
+    code = _http_status_of(err)
+    if code is not None:
+        return code == 429
     s = str(err or "").lower()
     if "429" in s or "too many requests" in s:
         return True
@@ -1555,6 +1585,15 @@ def _is_backend_unavailable(err) -> bool:
     # down". Checked FIRST because a 429 body can also contain the word "timeout".
     if _is_backend_overloaded(err):
         return False
+    # A structured gateway/timeout status IS an outage regardless of wording: a provider
+    # that returns `Error code: 502 - {"error": …}` carries no "bad gateway" text for the
+    # phrase list below to match, and would otherwise be misfiled as a generic error.
+    # 408 and Cloudflare's 522/524 are included so a real timeout STATUS keeps the same
+    # answer the "timed out" text below has always given it — without them, adding the
+    # structured check would have quietly demoted those.
+    code = _http_status_of(err)
+    if code is not None:
+        return code in (502, 503, 504, 408, 522, 524)
     return (
         "502 bad gateway" in s
         or "503 service" in s

package/package.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.326",
+  "version": "1.0.328",
   "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",