Updates

vendor/agent-harness/agent/compaction.js

@@ -24,7 +24,7 @@ const VOICE_COMPACT_PROMPT =
   '1. Latest user request\n' +
   '2. Facts from tools\n' +
   '3. What was already answered\n' +
-  '4. Open follow-ups\n' +
+  '4. Open follow-ups, original goal, user preferences and corrections\n' +
   'Be specific. Do not greet. Do not say the conversation was compacted.';
 
 function contentChars(content) {
@@ -75,16 +75,19 @@ function estimateTokens(messages, tools) {
   return conversationTokens(messages) + toolTokens(tools);
 }
 
-function historyBudget(ctxSize, tools, attempt) {
+function historyBudget(ctxSize, tools, attempt, extraTokens = 0) {
   const cap = ctxSize > 0 ? Number(ctxSize) : 8192;
   const toolTok = toolTokens(tools);
   const reserve = Math.max(384, Math.floor(cap * (0.18 + (Number(attempt) || 0) * 0.08)));
-  return Math.max(240, Math.floor(cap * 0.72) - toolTok - reserve);
+  return Math.max(240, Math.floor(cap * 0.72) - toolTok - reserve - extraTokens);
 }
 
-function shouldCompact(messages, tools, ctxSize) {
-  if (nonSystemCount(messages) < MIN_COMPACT_MESSAGES) return false;
-  return conversationTokens(messages) > historyBudget(ctxSize, tools, 0);
+function shouldCompact(messages, tools, ctxSize, extraTokens = 0) {
+  const budget = historyBudget(ctxSize, tools, 0, extraTokens);
+  const tokens = conversationTokens(messages);
+  // Large first requests or tool results can overflow before four messages.
+  // A small greeting must not be discarded merely because schemas are large.
+  return tokens > budget && (nonSystemCount(messages) >= MIN_COMPACT_MESSAGES || conversationTokens((messages || []).filter((m) => m.role !== 'system')) > 1024);
 }
 
 function isOverflowError(err) {
@@ -129,29 +132,31 @@ function isRealUser(m) {
   return true;
 }
 
+// Keep native tool exchanges atomic: results must never outlive their calls.
+function messageGroups(messages) {
+  const groups = [];
+  for (const message of messages) {
+    const previous = groups[groups.length - 1];
+    if (message.role === 'tool') {
+      const calls = previous && previous[0].tool_calls;
+      if (calls && calls.some((call) => call.id === message.tool_call_id)) previous.push(message);
+      continue;
+    }
+    groups.push([message]);
+  }
+  return groups;
+}
+
 function rebuildHistory(messages, summary) {
   const list = Array.isArray(messages) ? messages : [];
-  const sys =
-    list[0] && list[0].role === 'system'
-      ? { role: 'system', content: list[0].content }
-      : { role: 'system', content: '' };
-  let lastUser = null;
-  for (let i = list.length - 1; i >= 0; i--) {
-    if (isRealUser(list[i])) {
-      lastUser = { role: 'user', content: list[i].content };
-      break;
-    }
-  }
-  const tail = [];
-  for (let i = list.length - 1; i >= 1 && tail.length < 6; i--) {
-    const m = list[i];
-    if (!m || m.role === 'system') continue;
-    if (lastUser && m.role === 'user' && m.content === lastUser.content) continue;
-    tail.unshift({ role: m.role, content: m.content, name: m.name, tool_call_id: m.tool_call_id });
-  }
-  const out = [sys, { role: 'user', content: '[conversation summary]\n' + String(summary || '').trim() }];
-  if (lastUser) out.push(lastUser);
-  return out.concat(tail);
+  const systems = list.filter((m) => m.role === 'system');
+  const lastUser = lastRealUserIndex(list);
+  // Older turns belong in the summary, never after the latest request.
+  const tail = lastUser >= 0 ? list.slice(lastUser) : list.filter((m) => m.role !== 'system');
+  return systems.concat(
+    { role: 'user', content: '[conversation summary]\n' + String(summary || '').trim() },
+    messageGroups(tail.filter((m) => m.role !== 'system')).flat()
+  );
 }
 
 function truncateMsg(m, maxChars) {
@@ -186,44 +191,55 @@ function overHistoryBudget(keep, budget) {
 function heuristicCompact(messages, opts) {
   opts = opts || {};
   const budget = opts.budgetTokens || 6000;
-  const aggressive = !!opts.aggressive;
-  if (!aggressive && nonSystemCount(messages) < MIN_COMPACT_MESSAGES) {
-    return (messages || []).slice();
+  const list = messages || [];
+  const tokens = conversationTokens(list);
+  if (tokens <= budget || (!opts.aggressive && nonSystemCount(list) < MIN_COMPACT_MESSAGES && conversationTokens(list.filter((m) => m.role !== 'system')) <= 1024))
+    return list.slice();
+
+  const systems = list.filter((m) => m.role === 'system');
+  const lastUser = lastRealUserIndex(list);
+  const request = lastUser >= 0 ? list[lastUser] : null;
+  const older = list.slice(0, lastUser >= 0 ? lastUser : 0).filter((m) => m.role !== 'system');
+  const tail = list.slice(lastUser >= 0 ? lastUser + 1 : 0).filter((m) => m.role !== 'system');
+  let groups = messageGroups(tail);
+  // Preserve an existing summary across repeated compactions. The fallback
+  // carries real excerpts rather than replacing memory with a generic notice.
+  const memory = transcript(older.concat(tail));
+  const fixed = systems.concat(request ? [request] : []);
+  const spareChars = Math.max(0, (budget - conversationTokens(fixed) - 12) * CHAR_PER_TOKEN);
+  const summaryCap = Math.min(2400, Math.floor(spareChars * (groups.length ? 0.4 : 1)));
+  let summary = memory && summaryCap >= 64
+    ? { role: 'user', content: '[conversation summary]\n' + boundedExcerpt(memory, summaryCap) }
+    : null;
+  const assemble = () => systems.concat(summary ? [summary] : [], request ? [request] : [], groups.flat());
+  let keep = assemble();
+  // Shrink bulky results first; preserve call ids, arguments and ordering.
+  for (const cap of [1600, 600, 180]) {
+    if (!overHistoryBudget(keep, budget)) break;
+    groups = groups.map((group) => group.map((m) => truncateMsg(m, cap)));
+    keep = assemble();
   }
-  let keep = messages.slice();
-  while (overHistoryBudget(keep, budget) && keep.length > 4) {
-    let idx = keep.findIndex((m, i) => i > 0 && m.role === 'tool');
-    if (idx < 0) idx = keep.findIndex((m, i) => i > 1 && m.role === 'assistant');
-    if (idx < 0) break;
-    keep.splice(idx, 1);
+  while (overHistoryBudget(keep, budget) && groups.length) {
+    groups.shift();
+    keep = assemble();
   }
-  const maxMsg = aggressive ? 1200 : 3200;
-  if (overHistoryBudget(keep, budget)) {
-    const lastUser = lastRealUserIndex(keep);
-    keep = keep.map((m, i) => (i === 0 || i === lastUser ? m : truncateMsg(m, maxMsg)));
-  }
-  if (overHistoryBudget(keep, budget) && keep.length > 3) {
-    const head = keep[0];
-    const lastUserIdx = lastRealUserIndex(keep);
-    const lastUser = lastUserIdx >= 0 ? keep[lastUserIdx] : null;
-    const tail = [];
-    for (let i = keep.length - 1; i >= 1 && tail.length < (aggressive ? 2 : 4); i--) {
-      if (i === lastUserIdx) continue;
-      tail.unshift(keep[i]);
-    }
-    keep = [head, { role: 'user', content: '[conversation summary]\nEarlier turns were compacted to fit the model window.' }];
-    if (lastUser) keep.push(lastUser);
-    keep = keep.concat(tail);
-    keep = keep.map((m, i) => (i === 0 ? m : truncateMsg(m, aggressive ? 700 : 1800)));
-  }
-  while (overHistoryBudget(keep, budget) && keep.length > 3) {
-    const dropAt = keep.findIndex((m, i) => i > 1 && !isRealUser(m));
-    if (dropAt < 0) break;
-    keep.splice(dropAt, 1);
+  if (overHistoryBudget(keep, budget) && summary) {
+    summary = null;
+    keep = assemble();
   }
+  // The system instructions and latest request are never silently truncated.
+  // If these alone exceed the window, the caller reports the overflow.
   return keep;
 }
 
+function boundedExcerpt(text, maxChars) {
+  if (text.length <= maxChars) return text;
+  const marker = '\n[earlier details omitted]\n';
+  const remaining = Math.max(0, maxChars - marker.length);
+  const head = Math.floor(remaining * 0.6);
+  return text.slice(0, head) + marker + text.slice(-(remaining - head));
+}
+
 function compact(messages, opts) {
   return heuristicCompact(messages, opts);
 }
@@ -233,7 +249,10 @@ function transcript(messages) {
     .map((m) => {
       const role = m && m.role ? m.role : 'unknown';
       const name = m && m.name ? ' ' + m.name : '';
-      return role + name + ':\n' + String(m && m.content != null ? m.content : '');
+      const content = m && m.content != null ? m.content : '';
+      const body = typeof content === 'string' ? content : JSON.stringify(content);
+      if (body.startsWith('[conversation summary]')) return body.slice('[conversation summary]'.length).trim();
+      return role + name + ':\n' + body + (m && m.tool_calls ? '\nTool calls: ' + JSON.stringify(m.tool_calls) : '');
     })
     .join('\n\n');
 }
@@ -275,7 +294,7 @@ function compactReminder(opts) {
 async function compactWithLlm(messages, opts) {
   opts = opts || {};
   const fallback = () => heuristicCompact(messages, opts);
-  if (!opts.aggressive && nonSystemCount(messages) < MIN_COMPACT_MESSAGES) {
+  if (!opts.aggressive && nonSystemCount(messages) < MIN_COMPACT_MESSAGES && conversationTokens((messages || []).filter((m) => m.role !== 'system')) <= 1024) {
     return (messages || []).slice();
   }
   const complete = opts.complete;
@@ -284,7 +303,7 @@ async function compactWithLlm(messages, opts) {
   const cap =
     opts.maxTranscriptChars ||
     Math.max(1500, Math.min(24000, Math.floor((opts.budgetTokens || 4000) * CHAR_PER_TOKEN * 0.45)));
-  const body = truncate.truncateWithMarker(transcript(messages), cap);
+  const body = truncate.truncateWithMarker(transcript((messages || []).filter((m) => m.role !== 'system')), cap);
   const prompt = voice ? VOICE_COMPACT_PROMPT : COMPACT_PROMPT;
   const sys = voice
     ? 'Reply with the four summary sections only. No tools. Do not greet.'
@@ -299,7 +318,7 @@ async function compactWithLlm(messages, opts) {
     });
     const summary = result && result.text;
     if (isDegenerate(summary)) return fallback();
-    return rebuildHistory(messages, summary);
+    return heuristicCompact(rebuildHistory(messages, summary), Object.assign({}, opts, { aggressive: true }));
   } catch (_) {
     return fallback();
   }