@obtoai/agent-bridge 0.1.0-beta.16 → 0.1.0-beta.17

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@obtoai/agent-bridge",
-  "version": "0.1.0-beta.16",
+  "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

@@ -53,17 +53,23 @@ const readTail = (filePath, maxBytes = TAIL_READ_BYTES) => {
   }
 };
 
-// Pull the last N user/assistant turns from the tail, oldest-first. Used at
-// adopt time so the bridge thread shows actual prior context (not just a
-// one-line preview) — Phase 6.2.2. Skips injection-pattern user messages so
-// the history feels like real conversation, not platform noise.
-const RECENT_MESSAGE_BODY_MAX = 1500;
-const extractRecentMessages = (jsonlTail, n = 5) => {
+// 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 out = [];
-  for (let i = lines.length - 1; i >= 0 && out.length < n; i--) {
-    const line = lines[i].trim();
+  const parsed = [];
+  for (const lineRaw of lines) {
+    const line = lineRaw.trim();
     if (!line) continue;
     let obj;
     try { obj = JSON.parse(line); } catch (_) { continue; }
@@ -86,15 +92,37 @@ const extractRecentMessages = (jsonlTail, n = 5) => {
         .join(' ');
     }
     text = text.replace(/\s+/g, ' ').trim();
-    if (!text) continue;
-    if (role === 'user' && isInjectionMessage(text)) continue;
-    out.unshift({
-      role: role,
-      body: text.length > RECENT_MESSAGE_BODY_MAX ? text.slice(0, RECENT_MESSAGE_BODY_MAX) : text,
-      ts: obj.timestamp || null,
-    });
+    if (!text) continue; // skip pure tool_use / tool_result lines (no text body)
+    parsed.push({ role, text, ts: obj.timestamp || null });
   }
-  return out;
+
+  // 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