Updates

test/compaction.test.js

@@ -222,6 +222,35 @@ test('sidecar context is reserved in the history budget', () => {
   assert.equal(compaction.historyBudget(8192, [], 0) - compaction.historyBudget(8192, [], 0, 512), 512);
 });
 
+test('voice sessions keep a short spoken tail even on large model windows', () => {
+  assert.equal(compaction.voiceKeepTokens(4096), 1024);
+  assert.equal(compaction.voiceKeepTokens(16384), 1024);
+  assert.equal(compaction.voiceKeepTokens(32768), 1024);
+  assert.ok(compaction.voiceKeepTokens(2048) <= 573);
+  const sys = { role: 'system', content: 'You are Jarvis. '.repeat(40) };
+  const short = [
+    sys,
+    { role: 'assistant', content: 'Hello!' },
+    { role: 'user', content: 'What time is it?' },
+  ];
+  assert.equal(compaction.shouldCompact(short, [], 32768, 0, { voice: true }), false);
+  const long = [
+    sys,
+    { role: 'user', content: 'Remember nest.' },
+    { role: 'assistant', content: 'details '.repeat(2000) },
+    { role: 'user', content: 'And the hostname?' },
+    { role: 'assistant', content: 'Still nest. ' + 'padding '.repeat(200) },
+  ];
+  assert.equal(compaction.shouldCompact(long, [], 32768, 0, { voice: true }), true);
+  const sysTok = compaction.conversationTokens([sys]);
+  const budget = compaction.historyBudget(32768, [], 0, 0, { voice: true, systemTokens: sysTok });
+  assert.ok(budget <= sysTok + compaction.VOICE_KEEP_TOKENS);
+  assert.ok(budget < compaction.historyBudget(32768, [], 0));
+  const out = compaction.compact(long, { budgetTokens: budget, voice: true });
+  assert.ok(compaction.dialogTokens(out) <= compaction.VOICE_KEEP_TOKENS + 32);
+  assert.ok(out.some((m) => m.content === 'And the hostname?'));
+});
+
 for (const overflow of [false, true]) {
   test(`turn loop persists compaction and sends one-shot reminders through retries (overflow=${overflow})`, async (t) => {
     const { mkdtempSync, rmSync } = await import('node:fs');
@@ -260,13 +289,11 @@ for (const overflow of [false, true]) {
       { role: 'user', content: 'Remember the hostname.' },
       { role: 'assistant', content: 'I will remember nest.' },
     ]);
-    let summaries = 0;
     let completions = 0;
     const promptSizes = [];
     engine.complete = async ({ history }) => {
       if (history[0].content.startsWith('Reply with the four summary sections')) {
-        summaries++;
-        return { text: 'The computer is named nest. The user wants brief answers. We have checked memory and should continue answering follow-up questions about this computer.' };
+        throw new Error('voice turns must not spend a completion on LLM summarization');
       }
       completions++;
       promptSizes.push(compaction.conversationTokens(history));
@@ -285,7 +312,6 @@ for (const overflow of [false, true]) {
       emit: (_kind, event) => updates.push(event),
     });
     assert.equal(result.text, 'Its hostname is nest.');
-    assert.equal(summaries, 1);
     assert.equal(completions, overflow ? 2 : 1);
     if (overflow) assert.ok(promptSizes[1] < promptSizes[0], 'overflow recovery must actually shrink the prompt');
     assert.ok(updates.some((ev) => ev.type === 'compaction' && ev.status === 'done'));