@tiens.nguyen/gonext-local-worker 1.0.128 → 1.0.130

package/gonext-local-worker.mjs

@@ -1506,7 +1506,7 @@ async function runAgentChatJob(job) {
       codingBaseURL: payload?.codingBaseURL ?? "",
       codingModelId: payload?.codingModelId ?? "",
       tools: payload?.tools ?? ["http_request"],
-      maxSteps: payload?.maxSteps ?? 8,
+      maxSteps: payload?.maxSteps ?? 5,
       // Tool-invocation mode for the agent: "code" (default, python CodeAgent) or
       // "toolcall" (structured JSON tool calls). Passed through only; the API/web can
       // set payload.agentToolMode to A/B — default keeps current behavior.
@@ -1526,7 +1526,12 @@ async function runAgentChatJob(job) {
       emailFrom: payload?.emailFrom ?? "",
       emailAllowList: payload?.emailAllowList ?? "",
     });
-    const timeoutMs = 300_000; // 5 min max for an agent run
+    // 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;
 
     let inThink = false;
     let finalText = "";

package/gonext_agent_chat.py

@@ -11,7 +11,7 @@ Reads on stdin:
     "codingBaseURL": str,        # optional: dedicated coding/reasoning model for the
     "codingModelId": str,        #   CodeAgent's tool-use loop; empty = reuse agentModelId
     "tools": ["http_request"],   # v1: only http_request
-    "maxSteps": int              # multi-step ReAct budget; default 8
+    "maxSteps": int              # multi-step ReAct budget; default 5
   }
 
 Emits NDJSON lines on stdout:
@@ -66,13 +66,18 @@ def _http_request_impl(method, url, headers=None, body=None, timeout=25):
         return f"Error: {e}"
 
 
-def _html_to_text(html_text, limit=4000):
+def _html_to_text(html_text, limit=3000):
     """Strip HTML to readable plain text (zero-dep). Used by fetch_url so the weak
     model receives prose instead of raw tags it cannot parse."""
     import html as _html
     text = html_text or ""
     # Drop script/style/head/svg noise wholesale (incl. their content).
     text = re.sub(r"(?is)<(script|style|head|noscript|svg|template)\b.*?</\1>", " ", text)
+    # Drop page chrome wholesale too — nav menus, headers, footers, sidebars, forms.
+    # Without this, a Wikipedia fetch spends ~1KB of the budget on "Jump to content /
+    # Main menu / Donate / Create account…" before any article text, which both starves
+    # the model of real content and bloats the per-step context (OOM risk on local MLX).
+    text = re.sub(r"(?is)<(nav|header|footer|aside|form|button|menu)\b.*?</\1>", " ", text)
     # Turn block-ending tags into newlines so document structure survives stripping.
     text = re.sub(r"(?i)<(br|/p|/div|/li|/tr|/h[1-6]|/section|/article)\s*>", "\n", text)
     # Remove all remaining tags.
@@ -969,7 +974,7 @@ def run_agent_chat(cfg):
     # overridable via cfg.maxSteps. The provide_final_answer override below is now only
     # the exhaustion fallback (loop ends without final_answer).
     try:
-        max_steps = int(cfg.get("maxSteps") or 8)
+        max_steps = int(cfg.get("maxSteps") or 5)
     except (TypeError, ValueError):
         max_steps = 8
     if max_steps < 1:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.128",
+  "version": "1.0.130",
   "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",