Updates

vendor/agent-harness/agent/loop.js

@@ -72,6 +72,22 @@ function loadedCtxSize() {
   return (loaded && loaded.ctxSize) || 8192;
 }
 
+function systemTokens(history) {
+  if (history && history[0] && history[0].role === 'system') return compaction.conversationTokens([history[0]]);
+  return 0;
+}
+
+function historyCompactOpts(session, toolDefs, ctxSize, sidecarTokens, budget, attempt) {
+  return {
+    budgetTokens: compaction.historyBudget(ctxSize, toolDefs, attempt || 0, sidecarTokens, {
+      voice: !!budget.voice,
+      systemTokens: systemTokens(session.history),
+    }),
+    tools: toolDefs,
+    voice: !!budget.voice,
+  };
+}
+
 function toolResultCap(budget) {
   const ctx = loadedCtxSize();
   const voice = !!(budget && budget.voice);
@@ -608,22 +624,20 @@ async function runTurn(ctx) {
       const sidecarTokens = compaction.conversationTokens(turnSidecars) + 256;
       const beforeUsage = compaction.usage(session.history, toolDefs, ctxSize);
       emitLive(emit, session.id, jobId, Object.assign({ type: 'context' }, beforeUsage));
-      if (compaction.shouldCompact(session.history, toolDefs, ctxSize, sidecarTokens)) {
+      if (compaction.shouldCompact(session.history, toolDefs, ctxSize, sidecarTokens, { voice: !!budget.voice })) {
         emitUpdate(emit, session.id, jobId, {
           type: 'compaction',
           status: 'start',
-          method: 'llm',
+          method: budget.voice ? 'heuristic' : 'llm',
           used: beforeUsage.used,
           limit: beforeUsage.limit,
           pct: beforeUsage.pct,
           threshold: beforeUsage.threshold,
         });
-        const compactOpts = {
-          budgetTokens: compaction.historyBudget(ctxSize, toolDefs, 0, sidecarTokens),
-          tools: toolDefs,
-          voice: !!budget.voice,
-        };
-        session.history = await compaction.compactWithLlm(session.history, Object.assign({}, compactOpts, {
+        const compactOpts = historyCompactOpts(session, toolDefs, ctxSize, sidecarTokens, budget, 0);
+        session.history = budget.voice
+          ? compaction.compact(session.history, compactOpts)
+          : await compaction.compactWithLlm(session.history, Object.assign({}, compactOpts, {
           complete: (opts) => engine.complete(Object.assign({}, opts, {
             desktopVision: false,
             timeoutMs: budget.completeTimeoutMs,
@@ -643,7 +657,7 @@ async function runTurn(ctx) {
         emitUpdate(emit, session.id, jobId, {
           type: 'compaction',
           status: 'done',
-          method: 'llm',
+          method: budget.voice ? 'heuristic' : 'llm',
           used: afterUsage.used,
           limit: afterUsage.limit,
           pct: afterUsage.pct,
@@ -715,7 +729,7 @@ async function runTurn(ctx) {
             // The estimate can be lower than the model's tokenizer count.
             // Every overflow retry must shrink even an apparently small history.
             budgetTokens: Math.min(
-              compaction.historyBudget(ctxSize, toolDefs, overflowTry + 1, sidecarTokens),
+              historyCompactOpts(session, toolDefs, ctxSize, sidecarTokens, budget, overflowTry + 1).budgetTokens,
               Math.max(1, Math.floor(compaction.conversationTokens(session.history) * 0.85))
             ),
             tools: toolDefs,
@@ -863,7 +877,9 @@ async function runTurn(ctx) {
         }
         if (stopEarly) break;
       }
-      if (prepared.length) toolBudget.markToolRound(budget);
+      if (prepared.some((item) => item.name !== 'todo_write' && item.name !== 'update_goal')) {
+        toolBudget.markToolRound(budget);
+      }
       if (stopEarly) {
         return endTurn(emit, session, jobId, tracker, stopEarly);
       }