@tiens.nguyen/gonext-local-worker 1.0.134 → 1.0.138

package/gonext-local-worker.mjs

@@ -1526,12 +1526,14 @@ async function runAgentChatJob(job) {
       emailFrom: payload?.emailFrom ?? "",
       emailAllowList: payload?.emailAllowList ?? "",
     });
-    // 7 min max for an agent run: up to 5 ReAct steps on a local 14B with a cold
-    // prompt cache can legitimately exceed the old 5-min cap. Must stay BELOW the web
-    // client's hard WS deadline (480s) so our clean timeout error reaches the user
-    // instead of a generic WS timeout. On expiry the python child is SIGKILLed, which
-    // also shows up as a BrokenPipeError in the mlx_lm.server log (expected).
-    const timeoutMs = 420_000;
+    // 10 min max for an agent run: 5 ReAct steps on a 14B with a cold prompt cache —
+    // or a remote Ollama box cold-loading a big model — can exceed the old caps. Must
+    // stay BELOW the web client's hard WS deadline (660s) so our clean timeout error
+    // reaches the user instead of a generic WS timeout. On expiry the python child is
+    // SIGKILLed, which also shows up as a BrokenPipeError in the model server log
+    // (expected). Liveness during long model calls comes from the python heartbeat
+    // (a keepalive step chunk every 45s), not from this cap.
+    const timeoutMs = 600_000;
 
     let inThink = false;
     let finalText = "";

package/gonext_agent_chat.py

@@ -23,6 +23,7 @@ import contextlib
 import json
 import re
 import sys
+import threading
 import time
 import traceback
 import urllib.request
@@ -368,12 +369,16 @@ def _web_search_impl(query):
     )
 
 
-def _detect_model_id(base_url, api_key=""):
-    """Ask an OpenAI-compatible server which model it serves.
+class _AgentConfigError(RuntimeError):
+    """Deterministic setup error (e.g. wrong model name/URL). Not retryable and not
+    worth a plain-reply degrade — the user needs the message itself to fix Settings."""
+
+
+def _list_model_ids(base_url, api_key=""):
+    """Return the model ids an OpenAI-compatible server reports at {base_url}/models.
 
-    Queries GET {base_url}/models and returns the first reported model id.
-    `base_url` already ends with /v1. Returns "" on any failure so callers can
-    fall back. Used when the user supplies a coding-model URL but no model name.
+    `base_url` already ends with /v1. Returns [] on any failure so callers can
+    fall back.
     """
     url = base_url.rstrip("/") + "/models"
     headers = {"Accept": "application/json"}
@@ -384,14 +389,31 @@ def _detect_model_id(base_url, api_key=""):
         with urllib.request.urlopen(req, timeout=10, context=_ssl_context()) as resp:
             payload = json.loads(resp.read().decode("utf-8", errors="replace"))
     except Exception as e:  # noqa: BLE001
-        _log(f"model detect failed {url}: {e}")
-        return ""
+        _log(f"model list failed {url}: {e}")
+        return []
     data = payload.get("data") if isinstance(payload, dict) else None
-    if isinstance(data, list) and data:
-        first = data[0]
-        if isinstance(first, dict) and isinstance(first.get("id"), str):
-            return first["id"].strip()
-    return ""
+    if not isinstance(data, list):
+        return []
+    return [d["id"].strip() for d in data
+            if isinstance(d, dict) and isinstance(d.get("id"), str) and d["id"].strip()]
+
+
+def _detect_model_id(base_url, api_key=""):
+    """Ask an OpenAI-compatible server which model it serves (first reported id).
+
+    Returns "" on any failure so callers can fall back. Used when the user supplies
+    a coding-model URL but no model name.
+    """
+    ids = _list_model_ids(base_url, api_key)
+    if not ids:
+        return ""
+    if len(ids) > 1:
+        # Ollama servers list every pulled model; "first" is arbitrary and may be a
+        # huge/slow one (seen live: picked gemma4:31b over qwen3:14b).
+        _log(f"model detect: server hosts {len(ids)} models {ids} — using "
+             f"{ids[0]!r}. Set the coding model NAME in Settings to choose "
+             "a specific one.")
+    return ids[0]
 
 
 def _summarise_step(step_log):
@@ -1750,18 +1772,52 @@ def run_agent_chat(cfg):
             # A transient "Connection error" to the local MLX server (seen mid-loop when the
             # context grows large) otherwise aborts the whole run — retry a couple of times
             # with a short backoff before giving up.
+            # Heartbeat: while a model call is in flight the agent emits NOTHING, and
+            # the web's no-progress watchdog kills the chat. A big/cold model (e.g. a
+            # 31B on a remote Ollama box) can take minutes on prompt-eval before the
+            # first byte — emit a keepalive step every 45s until the call returns.
+            hb_stop = threading.Event()
+
+            def _heartbeat():
+                waited = 0
+                while not hb_stop.wait(45):
+                    waited += 45
+                    _emit({"type": "step",
+                           "text": f"Model is still working… ({waited}s — a large or "
+                                   "cold-loading model can take a few minutes)"})
+                    _log(f"heartbeat: model call in flight {waited}s")
+
+            hb = threading.Thread(target=_heartbeat, daemon=True)
+            hb.start()
             last_err = None
-            for attempt in range(3):
-                try:
-                    msg = super().generate(*args, **kwargs)
-                    break
-                except Exception as e:  # noqa: BLE001
-                    last_err = e
-                    if attempt < 2:
-                        _log(f"generate transient error (attempt {attempt + 1}/3): {e} — retrying")
-                        time.sleep(1.5 * (attempt + 1))
-            else:
-                raise last_err
+            try:
+                for attempt in range(3):
+                    try:
+                        msg = super().generate(*args, **kwargs)
+                        break
+                    except Exception as e:  # noqa: BLE001
+                        emsg = str(e)
+                        # A 404 "model not found" is a config error (wrong model NAME
+                        # for this server, e.g. an MLX-style name against Ollama) —
+                        # deterministic, so retrying is pointless. Fail fast with the
+                        # server's actual model list so the user can fix Settings.
+                        if "404" in emsg and "not found" in emsg.lower():
+                            ids = _list_model_ids(coding_base_url, agent_api_key)
+                            have = f" This server has: {', '.join(ids)}." if ids else ""
+                            raise _AgentConfigError(
+                                f"The agent coding model {coding_model_id!r} does not "
+                                f"exist on {coding_base_url}.{have} Set 'Agent coding "
+                                "model name' in Settings → Agent to one of these exact "
+                                "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:
+                    raise last_err
+            finally:
+                hb_stop.set()
             try:
                 content = getattr(msg, "content", None)
                 if content is None:
@@ -1896,6 +1952,12 @@ def run_agent_chat(cfg):
             model_id=coding_model_id,
             api_base=coding_base_url,
             api_key=agent_api_key,
+            # Cap each HTTP attempt at 240s so one hung request (remote server
+            # cold-loading a big model) fails into OUR retry loop instead of eating
+            # the whole worker budget; by the retry the model is usually loaded.
+            # max_retries=0: the openai client's own silent retries would otherwise
+            # stack multiplicatively with ours (3 × 3 attempts).
+            client_kwargs={"timeout": 240.0, "max_retries": 0},
         )
         agent_tools = [http_request, web_search, fetch_url, calculate,
                        get_current_datetime, create_pdf]
@@ -1929,6 +1991,23 @@ def run_agent_chat(cfg):
         _log(f"done (deterministic, no summarizer call): {len(final_text)} chars")
         _emit({"type": "final", "text": final_text})
     except Exception as e:  # noqa: BLE001
+        # Config errors (wrong coding-model name/URL) must reach the user VERBATIM —
+        # they contain the fix (the server's real model list). smolagents wraps model
+        # exceptions in AgentGenerationError, so walk the cause chain to find ours.
+        cfg_err = None
+        cur, depth = e, 0
+        while cur is not None and depth < 6:
+            if isinstance(cur, _AgentConfigError):
+                cfg_err = cur
+                break
+            cur, depth = (getattr(cur, "__cause__", None)
+                          or getattr(cur, "__context__", None)), depth + 1
+        if cfg_err is None and "Settings → Agent" in str(e):
+            cfg_err = e  # wrapped without a cause chain; message survived
+        if cfg_err is not None:
+            _log(f"agent config error (not degrading): {cfg_err}")
+            _emit({"type": "final", "text": f"⚠️ {cfg_err}"})
+            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.
         _log(f"agent error: {e} — degrading to plain reply")
@@ -1945,10 +2024,18 @@ def _log(text: str):
     _emit({"type": "log", "text": text})
 
 
+_EMIT_LOCK = threading.Lock()
+
+
 def _emit(obj):
-    """Write one NDJSON line to the real stdout and flush immediately."""
-    _REAL_STDOUT.write(json.dumps(obj) + "\n")
-    _REAL_STDOUT.flush()
+    """Write one NDJSON line to the real stdout and flush immediately.
+
+    Lock-guarded: the model-call heartbeat thread emits concurrently with the
+    main loop, and interleaved writes would corrupt the NDJSON stream.
+    """
+    with _EMIT_LOCK:
+        _REAL_STDOUT.write(json.dumps(obj) + "\n")
+        _REAL_STDOUT.flush()
 
 
 def main():

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.134",
+  "version": "1.0.138",
   "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",