titan-agent 5.0.2 → 5.0.3

package/dist/agent/agentLoop.js

@@ -12,6 +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 { scanForSecrets } from "../security/secretGuard.js";
 import { fullExfilScan } from "../security/exfilScan.js";
 import { runShellHooks } from "../hooks/shellHooks.js";
@@ -148,7 +149,24 @@ function detectToolUseIntent(userMessage) {
     // Requests to check real system state that REQUIRES a tool
     /\b(?:what is|what's|show me|get) (?:the )?(?:current|actual) (?:uptime|hostname|ip|path|directory|pwd|time|date|memory|disk)\b/,
     // Hunt Finding #17 (2026-04-14): added `[\s:]+` so "run: ls" matches too.
-    /\brun[\s:]+['"`]?(?:echo|ls|pwd|uptime|whoami|date|uname|cat|grep|find|node|npm|git|which|ps|df|free)\b/
+    /\brun[\s:]+['"`]?(?:echo|ls|pwd|uptime|whoami|date|uname|cat|grep|find|node|npm|git|which|ps|df|free)\b/,
+    // Widget / gallery tool requests — canvas chat explicitly asks for gallery_search/gallery_get
+    /\b(?:call|use|run)\b.*?\b(?:gallery_search|gallery_get|gallery_list)\b/,
+    /\b(?:create|add|make|build|spawn)\b.*?\b(?:widget|panel|canvas|gallery)\b/,
+    /\b(?:widget|gallery)\b.*?\b(?:template|search|find|get|fetch)\b/,
+    // New system widget intents (v5.0.2 "forgotten features" surface)
+    /\b(?:backup|snapshot|archive)\b/,
+    /\b(?:training|train|specialist|model)\b/,
+    /\b(?:recipe|playbook|workflow|jarvis)\b/,
+    /\b(?:vram|gpu|memory|nvidia)\b/,
+    /\b(?:team|member|role|permission|rbac)\b/,
+    /\b(?:cron|schedule|job|timer)\b/,
+    /\b(?:checkpoint|restore|save state)\b/,
+    /\b(?:organism|drive|safety|alert|guardrail)\b/,
+    /\b(?:fleet|node|route|mesh)\b/,
+    /\b(?:captcha|browser|form fill|web automation)\b/,
+    /\b(?:paperclip|sidecar|helper)\b/,
+    /\b(?:test|flaky|failing|coverage|eval)\b/
   ];
   return intentPatterns.some((p) => p.test(msg));
 }
@@ -404,6 +422,10 @@ async function runAgentLoop(ctx) {
   let modelSwitchCount = 0;
   let selfHealExhausted = false;
   const failedModels = /* @__PURE__ */ new Set();
+  const budgetConfig = getDefaultBudget(ctx.config);
+  if (budgetConfig.maxTokens > 0) {
+    initBudget(ctx.sessionId, budgetConfig);
+  }
   let noToolsRetryCount = 0;
   const MAX_MODEL_SWITCHES = 2;
   let pivotCount = 0;
@@ -618,6 +640,13 @@ 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;
+        }
         let response;
         const thinkStart = Date.now();
         const isChatRound0Think = round === 0 && phase === "think" && ctx.completionStrategy === "single-round";
@@ -678,6 +707,9 @@ Try a COMPLETELY DIFFERENT strategy. Do NOT repeat the same tools or approach.`
         result.promptTokens += promptTokens;
         result.completionTokens += completionTokens;
         const costCheck = recordTokenUsage(ctx.sessionId, activeModel, promptTokens, completionTokens);
+        if (budgetConfig.maxTokens > 0) {
+          recordUsage(ctx.sessionId, promptTokens, completionTokens);
+        }
         const callCost = calculateActualCost(activeModel, { prompt: promptTokens, completion: completionTokens });
         const sessionGoal = getSessionGoal(ctx.sessionId);
         recordSpend(ctx.agentId || "default", sessionGoal?.goalId, callCost);
@@ -909,7 +941,7 @@ Execute the next ACTION.` });
                 continue;
               }
             }
-            const finalText = stripToolJson(response.content || "");
+            let finalText = stripToolJson(response.content || "");
             const lastToolResult = result.toolCallDetails.length > 0 ? result.toolCallDetails[result.toolCallDetails.length - 1] : null;
             const userMsgLower = (ctx.message || "").toLowerCase();
             const wantsVerbatim = /\b(?:verbatim|exactly|precise|literal|as(?: |-)is|raw output|just the output)\b/.test(userMsgLower);
@@ -924,6 +956,29 @@ Execute the next ACTION.` });
                 break;
               }
             }
+            const galleryGetCalls = result.toolCallDetails.filter((d) => d.name === "gallery_get");
+            logger.info(COMPONENT, `[AutoWidgetGate] Checking ${galleryGetCalls.length} gallery_get call(s) for system widgets`);
+            for (const call of galleryGetCalls) {
+              try {
+                const snippet = call.resultSnippet || "";
+                logger.info(COMPONENT, `[AutoWidgetGate] gallery_get snippet: ${snippet.slice(0, 200)}`);
+                const parsed = JSON.parse(snippet);
+                if (parsed.source && typeof parsed.source === "string" && parsed.source.startsWith("system:")) {
+                  const gateText = `_____widget
+{ "name": "${parsed.name || "Widget"}", "format": "system", "source": "${parsed.source}", "w": ${parsed.defaultSize?.w || 6}, "h": ${parsed.defaultSize?.h || 6} }`;
+                  if (!finalText.includes("_____widget")) {
+                    finalText = finalText ? `${finalText}
+
+${gateText}` : gateText;
+                    logger.info(COMPONENT, `[AutoWidgetGate] Injected _____widget gate for ${parsed.source}`);
+                  }
+                } else {
+                  logger.info(COMPONENT, `[AutoWidgetGate] Skipped \u2014 source is ${parsed.source}`);
+                }
+              } catch (e) {
+                logger.info(COMPONENT, `[AutoWidgetGate] Parse error: ${e.message}`);
+              }
+            }
             result.content = finalText;
             setCachedResponse(smartMessages, activeModel, result.content);
             phase = "done";
@@ -1287,8 +1342,18 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
             }
           }
         } else {
-          phase = "respond";
-          logger.info(COMPONENT, `[ThinkAct] Tools executed \u2014 entering respond phase (tools will be stripped)`);
+          const discoveryTools = ["gallery_search", "tool_search"];
+          const lastWasDiscovery = pendingToolCalls.length === 1 && discoveryTools.includes(pendingToolCalls[0].function.name);
+          const resultHintsFollowUp = lastWasDiscovery && toolResults.some(
+            (r) => /call gallery_get/i.test(r.content) || /hint:.*?(call|use|try|fetch|get|search)/i.test(r.content)
+          );
+          if (resultHintsFollowUp && round < ctx.effectiveMaxRounds - 1) {
+            logger.info(COMPONENT, `[DiscoveryChain] ${pendingToolCalls[0].function.name} hinted follow-up \u2014 allowing another think round`);
+            phase = "think";
+          } else {
+            phase = "respond";
+            logger.info(COMPONENT, `[ThinkAct] Tools executed \u2014 entering respond phase (tools will be stripped)`);
+          }
         }
         break;
       }
@@ -1311,6 +1376,13 @@ 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;
+        }
         let smartMessages;
         if (ctx.voiceFastPath) {
           smartMessages = ctx.messages;
@@ -1326,7 +1398,7 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
         }
         const respondDirective = {
           role: "user",
-          content: '[System directive for this reply only] Write the final answer for the user. RULES: (1) Do NOT narrate what the user asked \u2014 they already know. (2) Do NOT describe your reasoning, thinking, or past tool attempts. (3) Do NOT start with "The user asked", "Let me", "Actually", "Looking at", "Wait" \u2014 start with the result. (4) Report outcomes as facts in 1-3 sentences. (5) No XML, no tool call blocks, no meta-commentary. Just the answer.'
+          content: '[System directive for this reply only] Write the final answer for the user. RULES: (1) Do NOT narrate what the user asked \u2014 they already know. (2) Do NOT describe your reasoning, thinking, or past tool attempts. (3) Do NOT start with "The user asked", "Let me", "Actually", "Looking at", "Wait" \u2014 start with the result. (4) Report outcomes as facts in 1-3 sentences. (5) No XML, no tool call blocks, no meta-commentary. Just the answer. (6) EXCEPTION: if you need to create or update a widget on the canvas, you MAY emit a _____react or _____widget gate \u2014 these are NOT meta-commentary, they are the actual deliverable.'
         };
         smartMessages = [...smartMessages, respondDirective];
         const thinkingMode = ctx.thinkingOverride || ctx.config.agent.thinkingMode || "off";
@@ -1383,6 +1455,9 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
           break;
         }
         const costCheck = recordTokenUsage(ctx.sessionId, activeModel, response.usage?.promptTokens || 0, response.usage?.completionTokens || 0);
+        if (budgetConfig.maxTokens > 0) {
+          recordUsage(ctx.sessionId, response.usage?.promptTokens || 0, response.usage?.completionTokens || 0);
+        }
         const callCost = calculateActualCost(activeModel, { prompt: response.usage?.promptTokens || 0, completion: response.usage?.completionTokens || 0 });
         const sessionGoal = getSessionGoal(ctx.sessionId);
         recordSpend(ctx.agentId || "default", sessionGoal?.goalId, callCost);
@@ -1500,6 +1575,9 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
     budgetExhausted: result.budgetExhausted
   });
   result.traceContext = { traceId: traceCtx.traceId, spanId: traceCtx.spanId };
+  if (budgetConfig.maxTokens > 0) {
+    cleanupBudget(ctx.sessionId);
+  }
   return result;
 }
 async function checkAndProcessInbox(agentId) {