titan-agent 5.3.2 → 5.4.1

package/dist/agent/agentLoop.js

@@ -26,6 +26,7 @@ async function chatWithTimeout(fn, label, timeoutMs = 3e5) {
 import { buildSmartContext } from "./contextManager.js";
 import { isKilled } from "../safety/killSwitch.js";
 import { estimateTokens } from "../utils/tokens.js";
+import { DEFAULT_MAX_TOKENS } from "../utils/constants.js";
 import { getCachedResponse, setCachedResponse } from "./responseCache.js";
 import { shouldReflect, reflect, resetProgress, recordProgress, setProgressSession } from "./reflection.js";
 import { recordToolResult, classifyTaskType, recordToolPreference, getErrorResolution, recordErrorResolution } from "../memory/learning.js";
@@ -442,7 +443,7 @@ async function runAgentLoop(ctx) {
   let budgetWarningSent = false;
   let pendingToolCalls = [];
   let pendingAssistantContent = "";
-  const tokenBudget = ctx.voiceFastPath ? 2e3 : ctx.config.agent.tokenBudget || 12e3;
+  const tokenBudget = ctx.voiceFastPath ? 2e3 : ctx.config.agent.tokenBudget || DEFAULT_MAX_TOKENS;
   globalSteerQueues.set(ctx.sessionId, ctx.steerQueue ?? []);
   const traceCtx = newTraceContext();
   initOtel();
@@ -663,6 +664,23 @@ Try a COMPLETELY DIFFERENT strategy. Do NOT repeat the same tools or approach.`
               ctx.streamCallbacks.onToolCall?.(chunk.toolCall.function.name, JSON.parse(chunk.toolCall.function.arguments || "{}"));
             } else if (chunk.type === "error") {
               logger.error(COMPONENT, `Stream error: ${chunk.error}`);
+            } else if (chunk.type === "retry") {
+              logger.warn(COMPONENT, `Stream retrying ${chunk.provider}/${chunk.model} (attempt ${chunk.attempt}/${chunk.maxRetries}, reason=${chunk.reason}, delay=${chunk.delayMs}ms)`);
+              ctx.streamCallbacks.onRetry?.({
+                attempt: chunk.attempt,
+                maxRetries: chunk.maxRetries,
+                reason: chunk.reason,
+                provider: chunk.provider,
+                model: chunk.model,
+                delayMs: chunk.delayMs
+              });
+            } else if (chunk.type === "failover") {
+              logger.warn(COMPONENT, `Stream failover from ${chunk.originalProvider}/${chunk.originalModel}${chunk.error ? ` (${chunk.error})` : ""}`);
+              ctx.streamCallbacks.onFailover?.({
+                originalProvider: chunk.originalProvider,
+                originalModel: chunk.originalModel,
+                error: chunk.error
+              });
             }
           }
           const estCompletionTokens = estimateTokens(streamContent + JSON.stringify(streamToolCalls));
@@ -1423,6 +1441,23 @@ Examples: use read_file instead of cat/head/tail/grep, and edit_file instead of
               ctx.streamCallbacks.onToken(chunk.content);
             } else if (chunk.type === "error") {
               logger.error(COMPONENT, `Stream error: ${chunk.error}`);
+            } else if (chunk.type === "retry") {
+              logger.warn(COMPONENT, `Respond-phase stream retrying ${chunk.provider}/${chunk.model} (attempt ${chunk.attempt}/${chunk.maxRetries}, reason=${chunk.reason})`);
+              ctx.streamCallbacks.onRetry?.({
+                attempt: chunk.attempt,
+                maxRetries: chunk.maxRetries,
+                reason: chunk.reason,
+                provider: chunk.provider,
+                model: chunk.model,
+                delayMs: chunk.delayMs
+              });
+            } else if (chunk.type === "failover") {
+              logger.warn(COMPONENT, `Respond-phase stream failover from ${chunk.originalProvider}/${chunk.originalModel}`);
+              ctx.streamCallbacks.onFailover?.({
+                originalProvider: chunk.originalProvider,
+                originalModel: chunk.originalModel,
+                error: chunk.error
+              });
             }
           }
           response = {