@obtoai/agent-bridge 0.1.0-beta.15 → 0.1.0-beta.17

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@obtoai/agent-bridge",
-  "version": "0.1.0-beta.15",
+  "version": "0.1.0-beta.17",
   "description": "Local consumer for the OBTO Agent Bridge. Receives bridge events over SSE and drives a coding agent (Claude Code or OpenAI Codex) on your machine.",
   "license": "Apache-2.0",
   "author": "OBTO Inc.",

package/src/external-scanner.js

@@ -53,6 +53,78 @@ const readTail = (filePath, maxBytes = TAIL_READ_BYTES) => {
   }
 };
 
+// Pull the last N user/assistant *logical turns* from the tail, oldest-first.
+// Phase 6.2.3 — beta.17 fix: Claude Code streams a single assistant reply as
+// MULTIPLE JSONL lines (text → tool_use → tool_result → text → …), each its
+// own JSONL record. The earlier extractor took one line per turn and
+// fragmented responses into single-shard previews. Now we walk forward,
+// parse every text-bearing user/assistant line, and **coalesce consecutive
+// same-role lines into one logical turn** before slicing the last N. Result:
+// a real conversation, not a single mid-sentence excerpt.
+const RECENT_MESSAGE_BODY_MAX = 3000;
+const RECENT_TURN_COUNT = 10;
+
+const extractRecentMessages = (jsonlTail, n = RECENT_TURN_COUNT) => {
+  if (!jsonlTail) return [];
+  const lines = jsonlTail.split(/\r?\n/);
+  const parsed = [];
+  for (const lineRaw of lines) {
+    const line = lineRaw.trim();
+    if (!line) continue;
+    let obj;
+    try { obj = JSON.parse(line); } catch (_) { continue; }
+    let role = null, raw = null;
+    if (obj && obj.message && (obj.message.role || obj.type)) {
+      role = obj.message.role || (obj.type === 'user' ? 'user' : 'assistant');
+      raw = obj.message.content;
+    } else if (obj && obj.role && (obj.content || obj.text)) {
+      role = obj.role;
+      raw = obj.content != null ? obj.content : obj.text;
+    }
+    if (role !== 'user' && role !== 'assistant') continue;
+    if (raw == null) continue;
+    let text = '';
+    if (typeof raw === 'string') text = raw;
+    else if (Array.isArray(raw)) {
+      text = raw
+        .filter((p) => p && (p.type === 'text' || typeof p.text === 'string'))
+        .map((p) => String(p.text || ''))
+        .join(' ');
+    }
+    text = text.replace(/\s+/g, ' ').trim();
+    if (!text) continue; // skip pure tool_use / tool_result lines (no text body)
+    parsed.push({ role, text, ts: obj.timestamp || null });
+  }
+
+  // Coalesce consecutive same-role lines (one logical turn split across JSONL
+  // records). This is the actual fix — without it, assistant turns with mid-
+  // reply tool calls fragment into the first text shard only.
+  const coalesced = [];
+  for (const p of parsed) {
+    const last = coalesced[coalesced.length - 1];
+    if (last && last.role === p.role) {
+      last.text = (last.text + ' ' + p.text).replace(/\s+/g, ' ').trim();
+      last.ts = p.ts || last.ts;
+    } else {
+      coalesced.push({ role: p.role, text: p.text, ts: p.ts });
+    }
+  }
+
+  // Filter user turns that are pure platform-injection noise. Assistant turns
+  // are never filtered — they're always real Claude/Codex output.
+  const filtered = coalesced.filter((m) => {
+    if (m.role !== 'user') return true;
+    return !isInjectionMessage(m.text);
+  });
+
+  const sliced = filtered.slice(-n);
+  return sliced.map((m) => ({
+    role: m.role,
+    body: m.text.length > RECENT_MESSAGE_BODY_MAX ? m.text.slice(0, RECENT_MESSAGE_BODY_MAX) : m.text,
+    ts: m.ts,
+  }));
+};
+
 // Claude Code writes an LLM-generated title as a `type: "ai-title"` JSONL
 // record near the end of each session file (this is the same title VSCode's
 // session list shows — "Analyze MongoDB MCP server architecture" style).
@@ -251,12 +323,14 @@ const scanClaude = () => {
       // have an ai-title record yet.
       let title = extractAiTitleFromTail(tail);
       if (!title) title = extractTitleFromFile(filePath);
+      const recentMessages = extractRecentMessages(tail);
       out.push({
         source: 'claude',
         sessionId,
         projectDir: entry,                         // raw encoded form
         projectName: decodeClaudeProjectDir(entry), // best-effort decoded
-        title: title,                              // first-user-message slice
+        title: title,                              // ai-title (Claude) or first-user-message fallback
+        recentMessages: recentMessages,            // last ≤5 user/assistant turns for adopt-time history
         lastActivityAt: stat.mtimeMs,
         lastMessagePreview: lastMsg ? lastMsg.preview : '',
         lastMessageAuthor: lastMsg ? lastMsg.author : null,
@@ -324,16 +398,15 @@ const scanCodex = () => {
 
           const tail = readTail(filePath);
           const lastMsg = extractLastMessage(tail);
-          // Codex rollouts don't carry an ai-title record (as of this writing),
-          // so we just use the first-user-message fallback. If they add one
-          // later, swap in extractAiTitleFromTail like the claude path does.
           const title = extractTitleFromFile(filePath);
+          const recentMessages = extractRecentMessages(tail);
           out.push({
             source: 'codex',
             sessionId,
             projectDir: projectDir || `${y}/${m}/${d}`,
             projectName: projectDir || null,
             title: title,
+            recentMessages: recentMessages,
             lastActivityAt: stat.mtimeMs,
             lastMessagePreview: lastMsg ? lastMsg.preview : '',
             lastMessageAuthor: lastMsg ? lastMsg.author : null,