aegiscode 5.2.31 → 5.2.32

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "aegiscode",
-  "version": "5.2.31",
+  "version": "5.2.32",
   "description": "AEGIS CLI — AI-powered coding assistant",
   "type": "module",
   "bin": {

package/scripts/repro-tokencount.ts

@@ -0,0 +1,45 @@
+/**
+ * Reproduce the interactive turn flow: ContextManager.addMessage for each
+ * user/assistant turn, then the post-turn recount from useCommandProcessor.
+ * Verify the stored token count tracks the actual conversation.
+ */
+import { ContextManager } from '../src/context/ContextManager.js';
+import { TokenCounter } from '../src/context/TokenCounter.js';
+
+async function main() {
+  const cm = new ContextManager({});
+  await cm.createSession();
+
+  // Simulate what useCommandProcessor does per turn, with realistic content
+  const turns = 8;
+  for (let i = 0; i < turns; i++) {
+    const userMsg = `Turn ${i}: please review the function at src/agent/Agent.ts, explain the auto-compact logic, and tell me whether the inter-turn truncation respects the reserved response budget. ${'Details: '.repeat(3)}${i}`;
+    const assistantMsg = `## Analysis (turn ${i})\n\nThe auto-compact path checks TokenCounter.estimateMessagesTokens first, then falls back to the exact BPE count when over threshold. The reserved response budget of 10k tokens is subtracted from maxTokens before truncation, so a long tool output cannot starve the final answer.\n\n| step | tokens |\n|------|--------|\n| pre-filter | ~${100 + i * 50} |\n| exact count | ${200 + i * 100} |\n\nThis is a moderately long assistant reply to simulate real conversation bulk across many turns. ${'words '.repeat(10)}end.`;
+
+    // 1. user message (ContextManager.addMessage)
+    await cm.addMessage('user', userMsg);
+
+    // 2. assistant message (ContextManager.addMessage)
+    await cm.addMessage('assistant', assistantMsg);
+
+    // 3. post-turn recount (useCommandProcessor lines 427-431)
+    const modelName = 'claude-sonnet-4-20250514';
+    const totalTokens = TokenCounter.countTokens(
+      cm.getMessages().map(m => ({ role: m.role as any, content: m.content })),
+      modelName
+    );
+    cm.updateTokenCount(totalTokens);
+
+    const stored = cm.getTokenCount();
+    const raw = cm.getContext()!.layers.conversation.messages.length;
+    console.log(`after turn ${i + 1}: store=${stored} tokens, msgs=${raw}`);
+  }
+
+  const final = cm.getTokenCount();
+  console.log('\n=== RESULT ===');
+  console.log('turns:', turns, '| final stored token count:', final);
+  console.log(final > 4000 ? 'PASS: /compact would proceed' : 'FAIL: /compact would refuse (stuck under 4000)');
+  await cm.cleanup();
+}
+
+main().catch(e => { console.error(e); process.exit(1); });