@manny-est/node-red-flowpilot 0.4.1 → 0.5.1

package/lib/provider-openai-compatible.js

@@ -194,7 +194,66 @@ async function chat(settings, messages, options) {
 // `data: [DONE]`). Calls onDelta(text) for each content fragment as it
 // arrives and resolves with { content } containing the full concatenated
 // text once the stream ends. Used for chat streaming only.
-function postStream(urlString, headers, body, timeoutMs, onDelta) {
+// Splits a streaming content string on <think>...</think> tags so that
+// reasoning text is routed to onReasoning and the actual response text to
+// onContent. Handles tags that arrive split across multiple chunks via a
+// small lookahead buffer. Handles both:
+//   - SGLang/Nemotron style: delta.reasoning_content (separate field)
+//   - llama.cpp/LocalAI style: <think>...</think> embedded in delta.content
+// The two paths converge at onReasoningDelta in postStream — callers see
+// one uniform reasoning callback regardless of provider format.
+function createThinkTagSplitter(onContent, onReasoning) {
+  let phase = "seek"; // "seek" | "think" | "content"
+  let buf = "";
+  const OPEN = "<think>";
+  const CLOSE = "</think>";
+
+  function push(text) {
+    buf += text;
+    while (buf.length > 0) {
+      if (phase === "seek") {
+        const idx = buf.indexOf(OPEN);
+        if (idx === -1) {
+          // No opening tag visible — emit everything except the last few bytes
+          // that might be a partial tag, buffer the rest.
+          const safe = buf.length > OPEN.length - 1 ? buf.length - (OPEN.length - 1) : 0;
+          if (safe > 0) { onContent(buf.slice(0, safe)); buf = buf.slice(safe); }
+          break;
+        }
+        if (idx > 0) { onContent(buf.slice(0, idx)); }
+        buf = buf.slice(idx + OPEN.length);
+        phase = "think";
+      } else if (phase === "think") {
+        const idx = buf.indexOf(CLOSE);
+        if (idx === -1) {
+          const safe = buf.length > CLOSE.length - 1 ? buf.length - (CLOSE.length - 1) : 0;
+          if (safe > 0) { onReasoning(buf.slice(0, safe)); buf = buf.slice(safe); }
+          break;
+        }
+        if (idx > 0) { onReasoning(buf.slice(0, idx)); }
+        buf = buf.slice(idx + CLOSE.length);
+        // Skip whitespace/newline immediately after </think>
+        const trimmed = buf.replace(/^\s+/, "");
+        buf = trimmed;
+        phase = "content";
+      } else {
+        onContent(buf);
+        buf = "";
+        break;
+      }
+    }
+  }
+
+  function finish() {
+    if (!buf) { return; }
+    if (phase === "think") { onReasoning(buf); } else { onContent(buf); }
+    buf = "";
+  }
+
+  return { push, finish };
+}
+
+function postStream(urlString, headers, body, timeoutMs, onDelta, onReasoningDelta) {
   return new Promise((resolve, reject) => {
     let url;
     try {
@@ -233,12 +292,22 @@ function postStream(urlString, headers, body, timeoutMs, onDelta) {
         return;
       }
 
-      let buf = "";
+      let sseBuf = "";
       let full = "";
+      // When onReasoningDelta is provided, intercept <think>...</think> from
+      // delta.content in addition to the dedicated delta.reasoning_content field
+      // (llama.cpp/LocalAI style vs SGLang/Nemotron style — both converge here).
+      const thinkSplitter = onReasoningDelta
+        ? createThinkTagSplitter(
+            function (c) { if (firstTokenAt === null) { firstTokenAt = Date.now(); } full += c; onDelta(c); },
+            onReasoningDelta
+          )
+        : null;
+
       res.on("data", (chunk) => {
-        buf += chunk;
-        const lines = buf.split("\n");
-        buf = lines.pop(); // keep the last (possibly partial) line for next time
+        sseBuf += chunk;
+        const lines = sseBuf.split("\n");
+        sseBuf = lines.pop(); // keep the last (possibly partial) line for next time
 
         lines.forEach((line) => {
           line = line.trim();
@@ -255,17 +324,29 @@ function postStream(urlString, headers, body, timeoutMs, onDelta) {
 
           if (evt && evt.usage) { usage = evt.usage; }
 
-          const delta = evt && evt.choices && evt.choices[0] && evt.choices[0].delta
-            ? evt.choices[0].delta.content
-            : "";
-          if (delta) {
-            if (firstTokenAt === null) { firstTokenAt = Date.now(); }
-            full += delta;
-            onDelta(delta);
+          const deltaObj = evt && evt.choices && evt.choices[0] && evt.choices[0].delta;
+          if (deltaObj) {
+            if (deltaObj.content) {
+              if (thinkSplitter) {
+                thinkSplitter.push(deltaObj.content);
+              } else {
+                if (firstTokenAt === null) { firstTokenAt = Date.now(); }
+                full += deltaObj.content;
+                onDelta(deltaObj.content);
+              }
+            }
+            // Separate reasoning_content field (SGLang/Nemotron style). Both
+            // paths run in parallel — they are mutually exclusive per provider
+            // (llama.cpp puts reasoning in delta.content via <think> tags;
+            // SGLang puts it in delta.reasoning_content with delta.content "").
+            if (deltaObj.reasoning_content && onReasoningDelta) {
+              onReasoningDelta(deltaObj.reasoning_content);
+            }
           }
         });
       });
       res.on("end", () => {
+        if (thinkSplitter) { thinkSplitter.finish(); }
         const endedAt = Date.now();
         resolve({
           content: full,
@@ -287,7 +368,7 @@ function postStream(urlString, headers, body, timeoutMs, onDelta) {
   });
 }
 
-async function chatStream(settings, messages, onDelta) {
+async function chatStream(settings, messages, onDelta, onReasoningDelta) {
   const baseUrl = String(settings.baseUrl || "").replace(/\/+$/, "");
   if (!baseUrl) throw new Error("Provider base URL is required.");
   if (!settings.model) throw new Error("Model is required.");
@@ -309,7 +390,7 @@ async function chatStream(settings, messages, onDelta) {
     temperature,
     stream: true,
     stream_options: { include_usage: true }
-  }, settings.requestTimeoutMs || 180000, onDelta);
+  }, settings.requestTimeoutMs || 180000, onDelta, onReasoningDelta);
 
   return {
     content: result.content || "",
@@ -377,4 +458,20 @@ async function probeTools(settings) {
   }
 }
 
-module.exports = { chat, chatStream, probeTools, listModels };
+// ---------------------------------------------------------------------
+// Checks whether the provider's response includes reasoning_content —
+// the separate chain-of-thought field emitted by reasoning models
+// (e.g. Nemotron, DeepSeek-R1, o1-style). Reuses the connectivity test
+// response already obtained by the /test route rather than making a
+// second round-trip: caller passes the raw response object.
+// Returns { isReasoningModel: boolean }.
+// ---------------------------------------------------------------------
+function detectReasoning(rawResponse) {
+  const message = rawResponse && rawResponse.choices && rawResponse.choices[0]
+    && rawResponse.choices[0].message;
+  const isReasoningModel = !!(message && message.reasoning_content != null
+    && message.reasoning_content !== "");
+  return { isReasoningModel };
+}
+
+module.exports = { chat, chatStream, probeTools, listModels, detectReasoning };

package/lib/storage.js

@@ -58,6 +58,10 @@ function createStorage(userDir) {
     // own hardcoded AGENT_LOOP_MAX_STEPS (a different bound, for a
     // different loop).
     agentLoopMaxIterations: 5,
+    // When true, the build loop pauses at the "attach → review" transition
+    // and shows a checkpoint question ("Continue with AI review, or stop?")
+    // instead of auto-advancing. Default false = original auto-advance behavior.
+    loopHoldStep: false,
     // Lets the user silence the recurring secrets/size reminder bar after
     // typing an explicit acknowledgement in settings.
     suppressContextWarnings: false,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@manny-est/node-red-flowpilot",
-  "version": "0.4.1",
+  "version": "0.5.1",
   "description": "FlowPilot for Node-RED - an AI-powered development assistant sidebar",
   "main": "flowpilot.js",
   "keywords": [
@@ -14,7 +14,7 @@
   "license": "MIT",
   "repository": {
     "type": "git",
-    "url": "https://github.com/manny-est/flowpilot.git"
+    "url": "git+https://github.com/manny-est/flowpilot.git"
   },
   "homepage": "https://github.com/manny-est/flowpilot#readme",
   "bugs": {
@@ -38,7 +38,6 @@
   "files": [
     "flowpilot.js",
     "flowpilot.html",
-    "flowpilot-core.js",
     "flowpilot-core.css",
     "lib",
     "icons",