@tiens.nguyen/gonext-local-worker 1.0.149 → 1.0.150

package/gonext-local-worker.mjs

@@ -139,6 +139,10 @@ const OCR_TIMEOUT_MS = 180_000;
 const OCR_DEBUG = String(process.env.GONEXT_OCR_DEBUG ?? "").trim() === "1";
 const OCR_SKIP_CORRECTION = String(process.env.GONEXT_OCR_SKIP_CORRECTION ?? "").trim() === "1";
 const OCR_CORRECT_TIMEOUT_MS = 120_000;
+// Fire-and-forget warm-up of the agent's Ollama coding model. Generous because a
+// cold load of a big/multimodal model into VRAM can take ~20s+; the request isn't
+// awaited, so a long ceiling just avoids aborting an in-progress load.
+const AGENT_WARMUP_TIMEOUT_MS = 180_000;
 // OCR translation/correction is hardcoded to a managed Ollama service + model.
 // (The user-configurable "OCR translation model" chooser was removed.)
 const OCR_TRANSLATE_OLLAMA_URL = "https://ollama1.gomarsic.cc";
@@ -1432,9 +1436,47 @@ async function runTranscribeJob(job) {
   }
 }
 
+/**
+ * Fire-and-forget pre-load of the agent's Ollama coding model. Ollama evicts idle
+ * models (default OLLAMA_KEEP_ALIVE=5m), so the first agent step otherwise blocks
+ * ~20s+ while llama-server spins up and pages the weights into VRAM. The agent
+ * talks to Ollama over the OpenAI-compat /v1 endpoint, which IGNORES keep_alive —
+ * so we hit the NATIVE /api/generate here (no prompt = load-only, returns as soon
+ * as the model is resident) with keep_alive:-1 to trigger the load AND pin it in
+ * memory. Kicked off the moment the job is claimed so the load overlaps python
+ * startup + prompt assembly. Best-effort: a non-Ollama coding server just 404s and
+ * the failure is swallowed — the job proceeds regardless.
+ */
+function warmUpOllamaCodingModel(codingBaseURL, codingModelId) {
+  if (!codingBaseURL || !codingModelId) return;
+  // codingBaseURL is normalized to …/v1 for OpenAI-compat; strip it to reach the
+  // Ollama root where /api/generate lives.
+  const root = String(codingBaseURL).replace(/\/+$/, "").replace(/\/v1$/i, "");
+  const url = `${root}/api/generate`;
+  fetch(url, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model: codingModelId, keep_alive: -1 }),
+    signal: AbortSignal.timeout(AGENT_WARMUP_TIMEOUT_MS),
+  })
+    .then((res) => {
+      console.log(
+        `[gonext-worker] agent warm-up ${res.ok ? "ok (model loading/resident)" : `HTTP ${res.status}`} → ${url} model=${codingModelId}`
+      );
+    })
+    .catch((err) => {
+      const reason =
+        err?.name === "TimeoutError" ? "still loading in background" : err?.message || "unreachable";
+      console.log(`[gonext-worker] agent warm-up skipped (${reason}) → ${url}`);
+    });
+}
+
 async function runAgentChatJob(job) {
   const { jobId, payload } = job;
   const start = Date.now();
+  // Kick the Ollama coding model's cold load off immediately (non-blocking) so it
+  // overlaps prompt assembly instead of stalling the first step. See helper above.
+  warmUpOllamaCodingModel(payload?.codingBaseURL ?? "", payload?.codingModelId ?? "");
   const runRes = await workerFetch(`/api/worker/jobs/${jobId}`, {
     method: "PATCH",
     body: JSON.stringify({ jobStatus: "running" }),
@@ -1507,6 +1549,9 @@ async function runAgentChatJob(job) {
       codingModelId: payload?.codingModelId ?? "",
       tools: payload?.tools ?? ["http_request"],
       maxSteps: payload?.maxSteps ?? 5,
+      // Max web_search + fetch_url calls the agent may make per task (user-configurable
+      // in web Settings → Agent). Default 10; past it the retrieval tools refuse.
+      researchBudget: payload?.researchBudget ?? 10,
       // 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.

package/gonext_agent_chat.py

@@ -1133,6 +1133,15 @@ def run_agent_chat(cfg):
     if max_steps < 1:
         max_steps = 8
 
+    # Max web_search + fetch_url calls per task (user-configurable in web Settings →
+    # Agent). Default 10; past it the retrieval tools refuse and steer to finish.
+    try:
+        research_budget = int(cfg.get("researchBudget") or 10)
+    except (TypeError, ValueError):
+        research_budget = 10
+    if research_budget < 1:
+        research_budget = 10
+
     _log(
         f"start model={agent_model_id!r} base={agent_base_url!r} "
         f"codeModel={coding_model_id!r} codeBase={coding_base_url!r} maxSteps={max_steps}"
@@ -1440,9 +1449,9 @@ def run_agent_chat(cfg):
         "clean one often does not exist. After 1-2 searches/fetches, COMPILE what you have and "
         "finish: if a PDF was asked for, call create_pdf(text=<your compiled text>, title=...); "
         "otherwise call final_answer. NEVER repeat a web_search/fetch_url you already ran.\n"
-        "- RESEARCH BUDGET — you may make at most 3 web_search/fetch_url calls TOTAL per task "
-        "(rewording a query still counts). Reserve your remaining steps for create_pdf / "
-        "final_answer.\n"
+        f"- RESEARCH BUDGET — you may make at most {research_budget} web_search/fetch_url calls "
+        "TOTAL per task (rewording a query still counts). Reserve your remaining steps for "
+        "create_pdf / final_answer.\n"
         "- Do NOT put final_answer outside the code block.\n\n"
     )
     # Lead with the TASK so the weak model anchors on what's actually being asked —
@@ -1486,10 +1495,10 @@ def run_agent_chat(cfg):
     # Hard research budget: at most N retrieval calls (web_search + fetch_url) per run.
     # Past it, retrieval tools refuse and steer to create_pdf/final_answer. Keeps steps
     # free to actually FINISH — the live failure mode was burning ALL steps on
-    # slightly-different searches and never producing the asked-for PDF. PDF tasks get a
-    # TIGHTER budget (2) so the model is steered to create_pdf a step sooner and has more
-    # steps left to actually emit that call itself (avoids leaning on the fallback).
-    _research = {"used": 0, "max": 2 if agent_pdf_requested else 3}
+    # slightly-different searches and never producing the asked-for PDF. The ceiling is
+    # user-configurable (web Settings → Agent, default 10) and applies to both PDF and
+    # plain tasks.
+    _research = {"used": 0, "max": research_budget}
 
     def _research_spend(tool_name: str) -> str:
         """Return '' if under budget (and spend 1), else a refusal steering to finish."""

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.149",
+  "version": "1.0.150",
   "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",