@tiens.nguyen/gonext-local-worker 1.0.49 → 1.0.51

package/README.md

@@ -2,6 +2,32 @@
 Run:
   GONEXT_API_BASE=... GONEXT_WORKER_KEY=... npx -y --package @tiens.nguyen/gonext-local-worker gonext-local-worker
 
+## Agent chat mode
+
+Select **Agent** (instead of Chat) under the composer in the web app. Your
+free-form prompt is handed to a [smolagents](https://github.com/huggingface/smolagents)
+agent running on your local MLX/Ollama model. The agent can call tools (v1:
+`http_request`) and streams its thinking steps + final answer directly into the
+chat thread.
+
+Requires smolagents in the worker's Python environment:
+
+```sh
+pip install smolagents certifi
+```
+
+- `certifi` supplies a trusted CA bundle so the agent's `http_request` tool can
+  verify HTTPS certificates on macOS (where Python's default bundle may be
+  missing). The worker falls back to the system bundle if certifi is absent.
+- Agent mode is blocked for cloud models (the API returns 400). Select a local
+  MLX or Ollama model first.
+- Tool steps appear in the collapsible reasoning (`<think>`) area; the final
+  answer is the message body — no new UI needed.
+
+The agent script is `gonext_agent_chat.py` (reads `{messages, agentBaseURL,
+agentApiKey, agentModelId, tools, maxSteps}` on stdin; emits NDJSON
+`{"type":"step"/"final","text":"..."}` lines on stdout).
+
 ## API Check / HTTP probe (Tools & Agents modes)
 
 The worker can run Postman-style HTTP probes queued from the web app
@@ -12,23 +38,15 @@ network_error).
 
 - **Tools (`tool_only`)** — no extra setup. The selected local model writes a
   one-line health summary of the measured result.
-- **Agents (`agentic`)** — a [smolagents](https://github.com/huggingface/smolagents)
-  agent (running on the selected local model) produces the summary. Install it
-  in the worker's Python environment:
-
-  ```sh
-  pip install smolagents
-  ```
+- **Agents (`agentic`)** — a smolagents agent (running on the selected local
+  model) produces the summary. Requires `pip install smolagents`.
 
   The agent talks to your local MLX OpenAI-compatible server (no cloud calls).
   The agent only summarizes; the worker's measurement stays the source of truth,
   so if smolagents or the model is unavailable the probe still returns the
   measured result with a note.
 
-### Probe-related env
+### Env vars
 
-  GONEXT_PROBE_PYTHON   Python executable for the smolagents agent
+  GONEXT_PROBE_PYTHON   Python executable for smolagents scripts
                         (default: GONEXT_MLX_LM_PYTHON or python3)
-
-The agent script lives next to the worker as `gonext_probe_agent.py` (reads a
-JSON probe config on stdin, writes a JSON summary on stdout).

package/gonext-local-worker.mjs

@@ -1216,6 +1216,16 @@ async function runAgentChatJob(job) {
     }
   };
 
+  console.log(
+    `[gonext-worker] agent_chat ${jobId} baseURL=${payload?.agentBaseURL ?? "(none)"} modelId=${payload?.agentModelId ?? "(none)"}`
+  );
+  // Send an immediate heartbeat so the web 60-180s no-progress timer doesn't
+  // fire while the local model is loading/generating its first reasoning step.
+  enqueueText("<think>Agent starting…\n");
+  flushTail = flushTail.then(() => flushChunks()).catch((err) => {
+    console.error("[gonext-worker] agent_chat heartbeat flush error:", err);
+  });
+
   try {
     const python =
       (process.env.GONEXT_PROBE_PYTHON ?? process.env.GONEXT_MLX_LM_PYTHON ?? "")
@@ -1231,7 +1241,7 @@ async function runAgentChatJob(job) {
     });
     const timeoutMs = 300_000; // 5 min max for an agent run
 
-    let inThink = false;
+    let inThink = true; // already opened <think> above
     let finalText = "";
 
     await runProcessWithStreamingStdout(python, [scriptPath], input, timeoutMs, (event) => {

package/gonext_agent_chat.py

@@ -27,6 +27,11 @@ import sys
 import urllib.request
 import urllib.error
 
+# Capture stdout before anything can redirect it.  _emit() must always write
+# to the real fd-1 so the Node worker's readline loop sees NDJSON even while
+# contextlib.redirect_stdout(sys.stderr) is active inside agent.run().
+_REAL_STDOUT = sys.stdout
+
 
 def _ssl_context():
     import ssl
@@ -143,6 +148,9 @@ def run_agent_chat(cfg):
             max_steps=max_steps,
             step_callbacks=[step_callback],
         )
+        # Emit before agent.run() so the web no-progress timer resets while the
+        # model loads its weights and generates its first reasoning step.
+        _emit({"type": "step", "text": f"Sending task to {agent_model_id}…"})
         with contextlib.redirect_stdout(sys.stderr):
             result = agent.run(task_text)
         final_text = str(result).strip()
@@ -152,9 +160,11 @@ def run_agent_chat(cfg):
 
 
 def _emit(obj):
-    """Write one NDJSON line to stdout and flush immediately."""
-    sys.stdout.write(json.dumps(obj) + "\n")
-    sys.stdout.flush()
+    """Write one NDJSON line to the real stdout and flush immediately.
+    Uses _REAL_STDOUT (captured at import time) so redirect_stdout blocks
+    inside agent.run() don't accidentally route events to stderr."""
+    _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": "1.0.49",
+  "version": "1.0.51",
   "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",