Updates

vendor/agent-harness/agent/compaction.js

@@ -1,5 +1,8 @@
 /**
  * Context compaction: heuristic fallback + optional one-shot LLM summary.
+ *
+ * Tool schemas are reserved out of historyBudget. Compact against conversation
+ * tokens only — never treat a large tool list as "history is full".
  */
 
 const truncate = require('./truncate.js');
@@ -7,6 +10,7 @@ const truncate = require('./truncate.js');
 const CHAR_PER_TOKEN = 3;
 const THRESHOLD = 0.68;
 const MIN_SUMMARY = 80;
+const MIN_COMPACT_MESSAGES = 4;
 const COMPACT_PROMPT =
   'Summarize this coding-agent conversation. Use exactly these sections:\n' +
   '1. Goal\n' +
@@ -15,6 +19,13 @@ const COMPACT_PROMPT =
   '4. Open work\n' +
   '5. Next action\n' +
   'Be specific (paths, names, errors). Do not say the conversation was compacted.';
+const VOICE_COMPACT_PROMPT =
+  'Summarize this spoken assistant conversation for the next turn. Use exactly these sections:\n' +
+  '1. Latest user request\n' +
+  '2. Facts from tools\n' +
+  '3. What was already answered\n' +
+  '4. Open follow-ups\n' +
+  'Be specific. Do not greet. Do not say the conversation was compacted.';
 
 function contentChars(content) {
   if (content == null) return 0;
@@ -42,23 +53,38 @@ function messageChars(m) {
   return n + 8;
 }
 
-function estimateTokens(messages, tools) {
+function nonSystemCount(messages) {
+  let n = 0;
+  for (const m of messages || []) {
+    if (m && m.role !== 'system') n += 1;
+  }
+  return n;
+}
+
+function conversationTokens(messages) {
   let n = 0;
   for (const m of messages || []) n += messageChars(m);
-  n += JSON.stringify(tools || []).length;
   return Math.ceil(n / CHAR_PER_TOKEN);
 }
 
+function toolTokens(tools) {
+  return Math.ceil(JSON.stringify(tools || []).length / CHAR_PER_TOKEN);
+}
+
+function estimateTokens(messages, tools) {
+  return conversationTokens(messages) + toolTokens(tools);
+}
+
 function historyBudget(ctxSize, tools, attempt) {
   const cap = ctxSize > 0 ? Number(ctxSize) : 8192;
-  const toolTok = Math.ceil(JSON.stringify(tools || []).length / CHAR_PER_TOKEN);
+  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);
 }
 
 function shouldCompact(messages, tools, ctxSize) {
-  const cap = ctxSize > 0 ? ctxSize : 8192;
-  return estimateTokens(messages, tools) > Math.floor(cap * THRESHOLD);
+  if (nonSystemCount(messages) < MIN_COMPACT_MESSAGES) return false;
+  return conversationTokens(messages) > historyBudget(ctxSize, tools, 0);
 }
 
 function isOverflowError(err) {
@@ -153,23 +179,30 @@ function lastRealUserIndex(list) {
   return -1;
 }
 
+function overHistoryBudget(keep, budget) {
+  return conversationTokens(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();
+  }
   let keep = messages.slice();
-  while (estimateTokens(keep, opts.tools) > budget && keep.length > 4) {
+  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);
   }
   const maxMsg = aggressive ? 1200 : 3200;
-  if (estimateTokens(keep, opts.tools) > budget) {
+  if (overHistoryBudget(keep, budget)) {
     const lastUser = lastRealUserIndex(keep);
     keep = keep.map((m, i) => (i === 0 || i === lastUser ? m : truncateMsg(m, maxMsg)));
   }
-  if (estimateTokens(keep, opts.tools) > budget && keep.length > 3) {
+  if (overHistoryBudget(keep, budget) && keep.length > 3) {
     const head = keep[0];
     const lastUserIdx = lastRealUserIndex(keep);
     const lastUser = lastUserIdx >= 0 ? keep[lastUserIdx] : null;
@@ -183,7 +216,7 @@ function heuristicCompact(messages, opts) {
     keep = keep.concat(tail);
     keep = keep.map((m, i) => (i === 0 ? m : truncateMsg(m, aggressive ? 700 : 1800)));
   }
-  while (estimateTokens(keep, opts.tools) > budget && keep.length > 3) {
+  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);
@@ -205,7 +238,12 @@ function transcript(messages) {
     .join('\n\n');
 }
 
-function autoContinue(messages) {
+function isVoice(opts) {
+  return !!(opts && opts.voice);
+}
+
+function autoContinue(messages, opts) {
+  if (isVoice(opts)) return null;
   const list = messages || [];
   const last = list[list.length - 1];
   if (!last) return null;
@@ -219,7 +257,14 @@ function autoContinue(messages) {
   return null;
 }
 
-function compactReminder() {
+function compactReminder(opts) {
+  if (isVoice(opts)) {
+    return (
+      '<system-reminder>\n' +
+      'Context was compacted. Trust the summary and the last user request. Answer that request. Do not greet again.\n' +
+      '</system-reminder>'
+    );
+  }
   return (
     '<system-reminder>\n' +
     'Context was compacted. Trust the summary and the last user request. Re-read files before further edits. Follow AGENTS.md if present.\n' +
@@ -230,17 +275,25 @@ function compactReminder() {
 async function compactWithLlm(messages, opts) {
   opts = opts || {};
   const fallback = () => heuristicCompact(messages, opts);
+  if (!opts.aggressive && nonSystemCount(messages) < MIN_COMPACT_MESSAGES) {
+    return (messages || []).slice();
+  }
   const complete = opts.complete;
   if (typeof complete !== 'function') return fallback();
+  const voice = isVoice(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 prompt = voice ? VOICE_COMPACT_PROMPT : COMPACT_PROMPT;
+  const sys = voice
+    ? 'Reply with the four summary sections only. No tools. Do not greet.'
+    : 'Reply with the five summary sections only. No tools.';
   try {
     const result = await complete({
       history: [
-        { role: 'system', content: 'Reply with the five summary sections only. No tools.' },
-        { role: 'user', content: COMPACT_PROMPT + '\n\n---\n\n' + body },
+        { role: 'system', content: sys },
+        { role: 'user', content: prompt + '\n\n---\n\n' + body },
       ],
       tools: [],
     });
@@ -256,10 +309,15 @@ module.exports = {
   CHAR_PER_TOKEN,
   THRESHOLD,
   MIN_SUMMARY,
+  MIN_COMPACT_MESSAGES,
   COMPACT_PROMPT,
+  VOICE_COMPACT_PROMPT,
+  conversationTokens,
+  toolTokens,
   estimateTokens,
   historyBudget,
   shouldCompact,
+  nonSystemCount,
   isOverflowError,
   usage,
   snapshot,