titan-agent 5.6.6 → 5.7.0

package/dist/agent/agentLoop.js

@@ -12,7 +12,7 @@ import { updateIssue, startRun, endRun, addIssueComment, recordSpend } from "./c
 import { recordTokenUsage, routeModel } from "./costOptimizer.js";
 import { calculateActualCost } from "./costEstimator.js";
 import { getSessionGoal } from "./autonomyContext.js";
-import { initBudget, checkBudget, recordUsage, markExceeded, cleanupBudget, getDefaultBudget } from "./promptBudget.js";
+import { initBudget, checkBudget, recordUsage, markExceeded, cleanupBudget, getDefaultBudget, resetBudgetUsage } from "./promptBudget.js";
 import { scanForSecrets } from "../security/secretGuard.js";
 import { fullExfilScan } from "../security/exfilScan.js";
 import { runShellHooks } from "../hooks/shellHooks.js";
@@ -644,12 +644,20 @@ Try a COMPLETELY DIFFERENT strategy. Do NOT repeat the same tools or approach.`
         if (round === 0 && phase === "think" && !ctx.isAutonomous && !ctx.taskEnforcementActive && ctx.activeTools.length > 0 && detectToolUseIntent(ctx.message || "")) {
           logger.info(COMPONENT, "[ExplicitIntent] User explicitly requested tool use \u2014 forcing tool_choice=required for round 1");
         }
-        const budgetMsg = budgetConfig.maxTokens > 0 ? checkBudget(ctx.sessionId, budgetConfig) : null;
-        if (budgetMsg) {
-          markExceeded(ctx.sessionId);
-          result.content = budgetMsg;
-          phase = "done";
-          break;
+        const budgetCheck = budgetConfig.maxTokens > 0 ? checkBudget(ctx.sessionId, budgetConfig) : null;
+        if (budgetCheck) {
+          if (budgetCheck.action === "compress") {
+            const before = ctx.messages.length;
+            const compressed = buildSmartContext(ctx.messages, Math.floor(budgetConfig.maxTokens * 0.6));
+            ctx.messages = compressed;
+            resetBudgetUsage(ctx.sessionId);
+            logger.info(COMPONENT, `[Budget] Compressed context: ${before} \u2192 ${compressed.length} messages, budget reset`);
+          } else {
+            markExceeded(ctx.sessionId);
+            result.content = `\u26A0\uFE0F ${budgetCheck.message}`;
+            phase = "done";
+            break;
+          }
         }
         let response;
         const thinkStart = Date.now();
@@ -1403,12 +1411,20 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
           break;
         }
         logger.info(COMPONENT, `Respond phase \u2014 calling LLM without tools to generate final answer`);
-        const respondBudgetMsg = budgetConfig.maxTokens > 0 ? checkBudget(ctx.sessionId, budgetConfig) : null;
-        if (respondBudgetMsg) {
-          markExceeded(ctx.sessionId);
-          result.content = respondBudgetMsg;
-          phase = "done";
-          break;
+        const respondBudgetCheck = budgetConfig.maxTokens > 0 ? checkBudget(ctx.sessionId, budgetConfig) : null;
+        if (respondBudgetCheck) {
+          if (respondBudgetCheck.action === "compress") {
+            const before = ctx.messages.length;
+            const compressed = buildSmartContext(ctx.messages, Math.floor(budgetConfig.maxTokens * 0.6));
+            ctx.messages = compressed;
+            resetBudgetUsage(ctx.sessionId);
+            logger.info(COMPONENT, `[Budget] Respond-phase compressed: ${before} \u2192 ${compressed.length} messages`);
+          } else {
+            markExceeded(ctx.sessionId);
+            result.content = `\u26A0\uFE0F ${respondBudgetCheck.message}`;
+            phase = "done";
+            break;
+          }
         }
         let smartMessages;
         if (ctx.voiceFastPath) {