openfox 2.0.116 → 2.0.118

package/dist/{chunk-4HV3PDEK.js → chunk-IDUXGSQV.js}

@@ -7,14 +7,18 @@ import {
 } from "./chunk-2FCLCUVR.js";
 import {
   TurnMetrics,
+  clearLLMFailure,
   consumeStreamGenerator,
   createChatDoneEvent,
   createMessageDoneEvent,
   createMessageStartEvent,
   createToolCallEvent,
   createToolResultEvent,
+  evaluateLLMRetry,
+  recordLLMFailure,
+  sleepThroughRetryBackoff,
   streamLLMPure
-} from "./chunk-EQ7DK2TG.js";
+} from "./chunk-NP2SV2CX.js";
 import {
   COMPACTION_PROMPT,
   buildBasePrompt,
@@ -67,19 +71,21 @@ import {
   extractSensitivePathsFromCommand,
   registerPathConfirmation,
   requestPathAccess
-} from "./chunk-QQVPUZQB.js";
+} from "./chunk-XZGQZ42D.js";
 import {
   getCurrentContextWindowId,
   getCurrentWindowMessageOptions,
   getEventStore
-} from "./chunk-TGVDUK4N.js";
+} from "./chunk-JOCR4CF4.js";
 import {
   createChatDoneMessage,
+  createChatLLMRetryFailedMessage,
+  createChatLLMRetryMessage,
   createChatMessageMessage,
   createChatMessageUpdatedMessage,
   createChatPathConfirmationMessage,
   createQueueStateMessage
-} from "./chunk-4GHLULHB.js";
+} from "./chunk-XAAEP5XM.js";
 import {
   AskUserInterrupt,
   askUserTool
@@ -1103,6 +1109,7 @@ function hasTopLevelStatementSeparator(command) {
       continue;
     }
     if (ch === ";" || ch === "\n") return true;
+    if ((ch === "&" || ch === "|") && command[i + 1] === ch) return true;
   }
   return false;
 }
@@ -1759,6 +1766,14 @@ function drainQueue(sessionManager, sessionId, append, onMessage) {
   return { messages: asapMessages, hasMessages: asapMessages.length > 0 };
 }
 
+// src/server/runner/types.ts
+var DEFAULT_LLM_RETRY_POLICY = {
+  backoffMs: [1e3, 5e3, 3e4],
+  minIntervalMs: 6e4,
+  maxDurationMs: 30 * 6e4,
+  maxAttempts: 40
+};
+
 // src/server/chat/agent-loop.ts
 function emitPartialDoneEvents(_sessionId, assistantMsgId, statsIdentity, mode, turnMetrics, append, agentType) {
   const stats = turnMetrics.buildStats(statsIdentity, mode);
@@ -1791,10 +1806,12 @@ function emitDoneAndBreak(assistantMsgId, segments, statsIdentity, mode, turnMet
 }
 var MAX_TRUNCATION_RETRIES = 3;
 var CONTINUE_PROMPT = "Continue your previous response. Do NOT repeat what you already wrote.";
+var CONTINUE_AFTER_STREAM_ERROR_PROMPT = "The LLM stream was interrupted mid-response. Continue exactly where you left off \u2014 do not repeat what was already written.";
 async function runTopLevelAgentLoop(config, turnMetrics) {
   const { mode, sessionManager, sessionId, llmClient, signal, onMessage, statsIdentity } = config;
   const append = config.append;
   const agentType = config.subAgentMetadata ? "sub-agent" : void 0;
+  const resolveClient = () => config.getLLMClient?.() ?? llmClient;
   const retryLimiter = createRetryLimiter(config.maxRetriesPerTurn ?? 10);
   let truncationRetryCount = 0;
   let returnValueContent;
@@ -1808,26 +1825,26 @@ async function runTopLevelAgentLoop(config, turnMetrics) {
     if (config.warmup) {
       const session2 = sessionManager.requireSession(sessionId);
       const runtimeConfig2 = getRuntimeConfig();
-      const configDir2 = getGlobalConfigDir(runtimeConfig2.mode ?? "production");
-      const skills2 = await getEnabledSkillMetadata(configDir2, sessionManager.getProjectWorkdir(sessionId));
+      const configDir = getGlobalConfigDir(runtimeConfig2.mode ?? "production");
+      const skills = await getEnabledSkillMetadata(configDir, sessionManager.getProjectWorkdir(sessionId));
       const { content: instructionContent2 } = await getAllInstructions(session2.workdir, session2.projectId);
       const toolRegistry2 = config.getToolRegistry();
-      const assembledRequest2 = await config.assembleRequest({
+      const assembledRequest = await config.assembleRequest({
         workdir: session2.workdir,
         messages: [],
         injectedFiles: [],
         promptTools: toolRegistry2.definitions,
         toolChoice: "none",
         ...instructionContent2 ? { customInstructions: instructionContent2 } : {},
-        ...skills2.length > 0 ? { skills: skills2 } : {}
+        ...skills.length > 0 ? { skills } : {}
       });
-      const modelSettings2 = sessionManager.getCurrentModelSettings(sessionId);
-      await llmClient.complete({
-        messages: [{ role: "system", content: assembledRequest2.systemPrompt }],
-        tools: assembledRequest2.tools,
+      const modelSettings = sessionManager.getCurrentModelSettings(sessionId);
+      await resolveClient().complete({
+        messages: [{ role: "system", content: assembledRequest.systemPrompt }],
+        tools: assembledRequest.tools,
         maxTokens: 1,
         temperature: 0,
-        ...modelSettings2 ? { modelSettings: modelSettings2 } : {}
+        ...modelSettings ? { modelSettings } : {}
       });
       return {};
     }
@@ -1844,71 +1861,125 @@ async function runTopLevelAgentLoop(config, turnMetrics) {
     }));
     const toolRegistry = config.getToolRegistry();
     const currentWindowMessageOptions = getCurrentWindowMessageOptions(sessionId);
-    const requestMessages = await config.getConversationMessages();
-    if (retryLimiter.count() > 0) {
-      const continueMsgId = crypto.randomUUID();
-      const continueContent = lastPatternMatch ? `Your previous response was interrupted because it matched pattern "${lastPatternMatch.pattern}" in ${lastPatternMatch.field}.
+    const retryPolicy = { ...DEFAULT_LLM_RETRY_POLICY, ...config.llmRetryPolicy };
+    const runtimeConfig = getRuntimeConfig();
+    let requestFailures = 0;
+    let requestFirstFailureAt = 0;
+    let continuationAppended = false;
+    let previousContextTokens;
+    let result;
+    let assistantMsgId;
+    let assistantMessageStarted = false;
+    for (; ; ) {
+      const requestMessages = await config.getConversationMessages();
+      if (requestFailures === 0 && retryLimiter.count() > 0) {
+        const continueMsgId = crypto.randomUUID();
+        const continueContent = lastPatternMatch ? `Your previous response was interrupted because it matched pattern "${lastPatternMatch.pattern}" in ${lastPatternMatch.field}.
 Matched content:
 ${lastPatternMatch.matchedContent}
 
 ${CONTINUE_PROMPT}` : CONTINUE_PROMPT;
-      append(
-        createMessageStartEvent(continueMsgId, "user", continueContent, {
-          ...currentWindowMessageOptions ?? {},
-          isSystemGenerated: true,
-          messageKind: "correction"
-        })
-      );
-      append({ type: "message.done", data: { messageId: continueMsgId } });
-      requestMessages.push({ role: "user", content: continueContent, source: "history" });
+        append(
+          createMessageStartEvent(continueMsgId, "user", continueContent, {
+            ...currentWindowMessageOptions ?? {},
+            isSystemGenerated: true,
+            messageKind: "correction"
+          })
+        );
+        append({ type: "message.done", data: { messageId: continueMsgId } });
+        requestMessages.push({ role: "user", content: continueContent, source: "history" });
+      }
+      const configDir = getGlobalConfigDir(runtimeConfig.mode ?? "production");
+      const skills = await getEnabledSkillMetadata(configDir, sessionManager.getProjectWorkdir(sessionId));
+      if (signal?.aborted) throw new Error("Aborted");
+      const assembledRequest = await config.assembleRequest({
+        workdir: session.workdir,
+        messages: requestMessages,
+        injectedFiles,
+        promptTools: toolRegistry.definitions,
+        toolChoice: "auto",
+        ...instructionContent ? { customInstructions: instructionContent } : {},
+        ...skills.length > 0 ? { skills } : {}
+      });
+      assistantMsgId = crypto.randomUUID();
+      assistantMessageStarted = false;
+      const ensureAssistantMessage = () => {
+        if (assistantMessageStarted) return;
+        assistantMessageStarted = true;
+        append(
+          createMessageStartEvent(assistantMsgId, "assistant", void 0, {
+            ...currentWindowMessageOptions ?? {},
+            ...config.subAgentMetadata ? { subAgentId: config.subAgentMetadata.subAgentId, subAgentType: config.subAgentMetadata.subAgentType } : {}
+          })
+        );
+      };
+      const contextState = sessionManager.getContextState(sessionId);
+      previousContextTokens = contextState.currentTokens;
+      const contextWindow = sessionManager.getCurrentModelContext();
+      const availableForOutput = Math.max(256, contextWindow - contextState.currentTokens);
+      let modelSettings = config.modelSettings ?? (currentMaxTokensOverride !== void 0 ? { ...sessionManager.getCurrentModelSettings(sessionId), maxTokens: currentMaxTokensOverride } : sessionManager.getCurrentModelSettings(sessionId));
+      if (modelSettings) {
+        const requestedMaxTokens = modelSettings.maxTokens ?? 16384;
+        modelSettings = { ...modelSettings, maxTokens: Math.min(requestedMaxTokens, availableForOutput) };
+      }
+      const allAgents = await loadAllAgentsDefault(sessionManager.getProjectWorkdir(sessionId));
+      const subAgentAliases = new Set(getSubAgents(allAgents).map((a) => a.metadata.id));
+      const streamGen = streamLLMPure({
+        messageId: assistantMsgId,
+        systemPrompt: assembledRequest.systemPrompt,
+        llmClient: resolveClient(),
+        messages: assembledRequest.messages,
+        tools: assembledRequest.tools,
+        toolChoice: "auto",
+        signal,
+        subAgentAliases,
+        ...config.retryPatterns ? { retryPatterns: config.retryPatterns } : {},
+        ...modelSettings && { modelSettings }
+      });
+      const attemptResult = await consumeStreamGenerator(streamGen, (event) => {
+        ensureAssistantMessage();
+        append(event);
+      });
+      if (!attemptResult.error) {
+        result = attemptResult;
+        break;
+      }
+      if (assistantMessageStarted && !continuationAppended) {
+        append(createMessageDoneEvent(assistantMsgId, { partial: true }));
+        onMessage?.(createChatMessageUpdatedMessage(assistantMsgId, { isStreaming: false, partial: true }));
+        const continueMsgId = crypto.randomUUID();
+        append(
+          createMessageStartEvent(continueMsgId, "user", CONTINUE_AFTER_STREAM_ERROR_PROMPT, {
+            ...currentWindowMessageOptions ?? {},
+            isSystemGenerated: true,
+            messageKind: "correction"
+          })
+        );
+        append({ type: "message.done", data: { messageId: continueMsgId } });
+        continuationAppended = true;
+      }
+      if (signal?.aborted) throw new Error("Aborted");
+      requestFailures += 1;
+      if (requestFirstFailureAt === 0) {
+        requestFirstFailureAt = Date.now();
+      }
+      const decision = evaluateLLMRetry(requestFailures, requestFirstFailureAt, Date.now(), retryPolicy);
+      if (!decision.retry) {
+        if (!config.subAgentMetadata) {
+          recordLLMFailure(sessionId);
+          config.onMessage?.(createChatLLMRetryFailedMessage(attemptResult.error, requestFailures));
+        }
+        return { failed: { error: attemptResult.error } };
+      }
+      if (!config.subAgentMetadata) {
+        config.onMessage?.(createChatLLMRetryMessage(decision.attempt, decision.delayMs));
+      }
+      const waitResult = await sleepThroughRetryBackoff(decision.delayMs, sessionId, signal);
+      if (waitResult === "aborted") throw new Error("Aborted");
+    }
+    if (!config.subAgentMetadata) {
+      clearLLMFailure(sessionId);
     }
-    const runtimeConfig = getRuntimeConfig();
-    const configDir = getGlobalConfigDir(runtimeConfig.mode ?? "production");
-    const skills = await getEnabledSkillMetadata(configDir, sessionManager.getProjectWorkdir(sessionId));
-    if (signal?.aborted) throw new Error("Aborted");
-    const assembledRequest = await config.assembleRequest({
-      workdir: session.workdir,
-      messages: requestMessages,
-      injectedFiles,
-      promptTools: toolRegistry.definitions,
-      toolChoice: "auto",
-      ...instructionContent ? { customInstructions: instructionContent } : {},
-      ...skills.length > 0 ? { skills } : {}
-    });
-    const assistantMsgId = crypto.randomUUID();
-    append(
-      createMessageStartEvent(assistantMsgId, "assistant", void 0, {
-        ...currentWindowMessageOptions ?? {},
-        ...config.subAgentMetadata ? { subAgentId: config.subAgentMetadata.subAgentId, subAgentType: config.subAgentMetadata.subAgentType } : {}
-      })
-    );
-    const contextState = sessionManager.getContextState(sessionId);
-    const previousContextTokens = contextState.currentTokens;
-    const contextWindow = sessionManager.getCurrentModelContext();
-    const availableForOutput = Math.max(256, contextWindow - contextState.currentTokens);
-    let modelSettings = config.modelSettings ?? (currentMaxTokensOverride !== void 0 ? { ...sessionManager.getCurrentModelSettings(sessionId), maxTokens: currentMaxTokensOverride } : sessionManager.getCurrentModelSettings(sessionId));
-    if (modelSettings) {
-      const requestedMaxTokens = modelSettings.maxTokens ?? 16384;
-      modelSettings = { ...modelSettings, maxTokens: Math.min(requestedMaxTokens, availableForOutput) };
-    }
-    const allAgents = await loadAllAgentsDefault(sessionManager.getProjectWorkdir(sessionId));
-    const subAgentAliases = new Set(getSubAgents(allAgents).map((a) => a.metadata.id));
-    const streamGen = streamLLMPure({
-      messageId: assistantMsgId,
-      systemPrompt: assembledRequest.systemPrompt,
-      llmClient,
-      messages: assembledRequest.messages,
-      tools: assembledRequest.tools,
-      toolChoice: "auto",
-      signal,
-      subAgentAliases,
-      ...config.retryPatterns ? { retryPatterns: config.retryPatterns } : {},
-      ...config.suppressChatError ? { suppressChatError: true } : {},
-      ...modelSettings && { modelSettings }
-    });
-    const result = await consumeStreamGenerator(streamGen, (event) => {
-      append(event);
-    });
     if (result.patternMatch) {
       if (!retryLimiter.canRetry()) {
         append({
@@ -1948,41 +2019,30 @@ ${CONTINUE_PROMPT}` : CONTINUE_PROMPT;
       continue;
     }
     if (result.aborted) {
-      emitPartialDoneEvents(sessionId, assistantMsgId, statsIdentity, mode, turnMetrics, append, agentType);
+      if (assistantMessageStarted) {
+        emitPartialDoneEvents(sessionId, assistantMsgId, statsIdentity, mode, turnMetrics, append, agentType);
+      }
       throw new Error("Aborted");
     }
-    if (result.error) {
-      const stats2 = turnMetrics.buildStats(statsIdentity, mode);
-      append(
-        createMessageDoneEvent(assistantMsgId, {
-          segments: result.segments,
-          stats: stats2
-        })
-      );
-      append(createChatDoneEvent(assistantMsgId, "error", stats2, agentType));
-      return { failed: { error: result.error } };
-    }
-    if (!result.error) {
-      turnMetrics.addLLMCall(
-        result.timing,
-        result.usage.promptTokens,
-        result.usage.completionTokens,
-        previousContextTokens,
-        result.modelParams
-      );
-      sessionManager.setCurrentContextSize(
-        sessionId,
-        result.usage.promptTokens,
-        result.usage.completionTokens,
-        config.subAgentMetadata?.subAgentId
-      );
-    }
+    turnMetrics.addLLMCall(
+      result.timing,
+      result.usage.promptTokens,
+      result.usage.completionTokens,
+      previousContextTokens,
+      result.modelParams
+    );
+    sessionManager.setCurrentContextSize(
+      sessionId,
+      result.usage.promptTokens,
+      result.usage.completionTokens,
+      config.subAgentMetadata?.subAgentId
+    );
     if (!compacting) {
-      const contextState2 = sessionManager.getContextState(sessionId);
-      const { shouldCompact, appendCompactionPrompt } = await import("./compactor-LAXZLPPA.js");
+      const contextState = sessionManager.getContextState(sessionId);
+      const { shouldCompact, appendCompactionPrompt } = await import("./compactor-5ATICQ2A.js");
       if (shouldCompact(
-        contextState2.currentTokens,
-        contextState2.maxTokens,
+        contextState.currentTokens,
+        contextState.maxTokens,
         sessionManager.getModelCompactionThreshold(sessionId) ?? runtimeConfig.context.compactionThreshold
       )) {
         appendCompactionPrompt(sessionId, append);
@@ -1995,8 +2055,8 @@ ${CONTINUE_PROMPT}` : CONTINUE_PROMPT;
         truncationRetryCount += 1;
         const currentMaxTokens = result.modelParams?.maxTokens ?? 16384;
         const promptTokens = result.usage.promptTokens;
-        const contextWindow2 = sessionManager.getCurrentModelContext();
-        const newMaxTokens = Math.min(Math.floor(currentMaxTokens * 1.5), contextWindow2 - promptTokens - 2048);
+        const contextWindow = sessionManager.getCurrentModelContext();
+        const newMaxTokens = Math.min(Math.floor(currentMaxTokens * 1.5), contextWindow - promptTokens - 2048);
         currentMaxTokensOverride = newMaxTokens;
         const interimStats = turnMetrics.buildStats(statsIdentity, mode);
         append(
@@ -2068,7 +2128,7 @@ ${COMPACTION_PROMPT}`,
           turnMetrics,
           signal,
           onMessage,
-          llmClient,
+          llmClient: resolveClient(),
           statsIdentity,
           onToolExecuted: config.onToolExecuted
         };
@@ -2129,7 +2189,9 @@ ${COMPACTION_PROMPT}`,
         emitPartialDoneEvents(sessionId, assistantMsgId, statsIdentity, mode, turnMetrics, append, agentType);
         throw new Error("Aborted");
       }
-      void drainQueue(sessionManager, sessionId, append, onMessage);
+      if (!config.subAgentMetadata) {
+        void drainQueue(sessionManager, sessionId, append, onMessage);
+      }
       retryLimiter.reset();
       continue;
     }
@@ -2967,7 +3029,7 @@ var callSubAgentTool = {
       };
     }
     try {
-      const { getToolRegistryForAgent: getToolRegistryForAgent2 } = await import("./tools-KBS6X3EN.js");
+      const { getToolRegistryForAgent: getToolRegistryForAgent2 } = await import("./tools-GOGLGTNV.js");
       const toolRegistry = getToolRegistryForAgent2(agentDef);
       const turnMetrics = new TurnMetrics();
       const result = await executeSubAgent({
@@ -3989,7 +4051,7 @@ var mcpConfigTool = createTool(
       mcpNotifyChanged?.(sessionId);
     }
     async function rebuildTools() {
-      const { setMcpTools: setMcpTools2 } = await import("./tools-KBS6X3EN.js");
+      const { setMcpTools: setMcpTools2 } = await import("./tools-GOGLGTNV.js");
       const mcpTools = createMcpTools(mcpManagerForTools);
       setMcpTools2(mcpTools);
     }
@@ -4873,4 +4935,4 @@ export {
   getToolRegistryForAgent,
   createToolRegistry
 };
-//# sourceMappingURL=chunk-4HV3PDEK.js.map
+//# sourceMappingURL=chunk-IDUXGSQV.js.map

package/dist/{chunk-TGVDUK4N.js → chunk-JOCR4CF4.js}

@@ -1465,4 +1465,4 @@ export {
   truncateSessionMessages,
   getRecentUserPromptsForSession
 };
-//# sourceMappingURL=chunk-TGVDUK4N.js.map
+//# sourceMappingURL=chunk-JOCR4CF4.js.map

package/dist/{chunk-ZKLE2Q3I.js → chunk-KSDWQ6IZ.js}

@@ -26,4 +26,4 @@ export {
   isClientMessage,
   isServerMessage
 };
-//# sourceMappingURL=chunk-ZKLE2Q3I.js.map
+//# sourceMappingURL=chunk-KSDWQ6IZ.js.map

package/dist/{chunk-6XHDNSA3.js → chunk-MTFLMNLN.js}

@@ -1,14 +1,14 @@
 import {
   combineEventsWithSnapshot,
   getEventStore
-} from "./chunk-TGVDUK4N.js";
+} from "./chunk-JOCR4CF4.js";
 import {
   updateSessionMetadata
 } from "./chunk-KLQEL2DZ.js";
 import {
   createContextStateMessage,
   createSessionStateMessage
-} from "./chunk-4GHLULHB.js";
+} from "./chunk-XAAEP5XM.js";
 import {
   getPendingQuestionsForSession
 } from "./chunk-JURM3RPZ.js";
@@ -51,6 +51,7 @@ function buildRunChatTurnParams(params) {
     sessionManager: params.sessionManager,
     sessionId: params.sessionId,
     llmClient: params.llmClient,
+    ...params.getSessionLLMClient ? { getSessionLLMClient: params.getSessionLLMClient } : {},
     signal: params.signal,
     onMessage: params.onMessage,
     ...params.statsIdentity ? { statsIdentity: params.statsIdentity } : {}
@@ -103,6 +104,7 @@ function applyGeneratedSessionName(sessionId, name, deps) {
     const { messages, hiddenCount } = buildMessagesFromStoredEvents(events, maxVisibleItems || void 0);
     const pendingConfirmations = foldPendingConfirmations(events);
     const pendingQuestions = getPendingQuestionsForSession(sessionId);
+    const activeWorkflowExecution = deps.sessionManager.getDisplayWorkflowExecution?.(sessionId) ?? void 0;
     deps.broadcastForSession(
       sessionId,
       createSessionStateMessage(
@@ -112,7 +114,8 @@ function applyGeneratedSessionName(sessionId, name, deps) {
         pendingQuestions,
         void 0,
         void 0,
-        hiddenCount
+        hiddenCount,
+        activeWorkflowExecution
       )
     );
   }
@@ -204,4 +207,4 @@ export {
   needsNameGeneration,
   generateSessionNameForSession
 };
-//# sourceMappingURL=chunk-6XHDNSA3.js.map
+//# sourceMappingURL=chunk-MTFLMNLN.js.map

package/dist/{chunk-EQ7DK2TG.js → chunk-NP2SV2CX.js}

@@ -244,11 +244,6 @@ async function* streamLLMPure(options) {
         case "error":
           if (signal?.aborted) break;
           streamError = value.error;
-          if (options.suppressChatError) break;
-          yield {
-            type: "chat.error",
-            data: { error: value.error, recoverable: true }
-          };
           break;
       }
       if (activePatterns.length > 0 && (accumulatedContent || accumulatedThinking)) {
@@ -291,6 +286,73 @@ async function* streamLLMPure(options) {
   }
   return baseResult;
 }
+function evaluateLLMRetry(failures, firstFailureAt, now, policy) {
+  if (now - firstFailureAt >= policy.maxDurationMs) return { retry: false };
+  if (failures >= policy.maxAttempts) return { retry: false };
+  const delayMs = failures <= policy.backoffMs.length ? policy.backoffMs[failures - 1] : policy.minIntervalMs;
+  return { retry: true, delayMs, attempt: failures + 1 };
+}
+var retryWaiters = /* @__PURE__ */ new Map();
+var llmFailures = /* @__PURE__ */ new Map();
+function hasRecentLLMFailure(sessionId, withinMs) {
+  const at = llmFailures.get(sessionId);
+  if (at === void 0) return false;
+  if (Date.now() - at <= withinMs) return true;
+  llmFailures.delete(sessionId);
+  return false;
+}
+function interruptLLMRetryWait(sessionId) {
+  const controller = retryWaiters.get(sessionId);
+  if (!controller) return false;
+  controller.abort();
+  return true;
+}
+function sleepWithRetryInterrupt(ms, signal, retryNowSignal) {
+  return new Promise((resolve) => {
+    if (signal?.aborted) {
+      resolve("aborted");
+      return;
+    }
+    if (retryNowSignal?.aborted) {
+      resolve("retry-now");
+      return;
+    }
+    const cleanup = () => {
+      clearTimeout(timer);
+      signal?.removeEventListener("abort", onAbort);
+      retryNowSignal?.removeEventListener("abort", onRetryNow);
+    };
+    const onAbort = () => {
+      cleanup();
+      resolve("aborted");
+    };
+    const onRetryNow = () => {
+      cleanup();
+      resolve("retry-now");
+    };
+    const timer = setTimeout(() => {
+      cleanup();
+      resolve("waited");
+    }, ms);
+    signal?.addEventListener("abort", onAbort, { once: true });
+    retryNowSignal?.addEventListener("abort", onRetryNow, { once: true });
+  });
+}
+async function sleepThroughRetryBackoff(ms, sessionId, signal) {
+  const retryNowController = new AbortController();
+  retryWaiters.set(sessionId, retryNowController);
+  try {
+    return await sleepWithRetryInterrupt(ms, signal, retryNowController.signal);
+  } finally {
+    retryWaiters.delete(sessionId);
+  }
+}
+function recordLLMFailure(sessionId) {
+  llmFailures.set(sessionId, Date.now());
+}
+function clearLLMFailure(sessionId) {
+  llmFailures.delete(sessionId);
+}
 var TurnMetrics = class {
   startTime;
   totalPrefillTokens = 0;
@@ -434,6 +496,12 @@ async function consumeStreamGenerator(gen, onEvent) {
 
 export {
   streamLLMPure,
+  evaluateLLMRetry,
+  hasRecentLLMFailure,
+  interruptLLMRetryWait,
+  sleepThroughRetryBackoff,
+  recordLLMFailure,
+  clearLLMFailure,
   TurnMetrics,
   createMessageStartEvent,
   createMessageDoneEvent,
@@ -442,4 +510,4 @@ export {
   createChatDoneEvent,
   consumeStreamGenerator
 };
-//# sourceMappingURL=chunk-EQ7DK2TG.js.map
+//# sourceMappingURL=chunk-NP2SV2CX.js.map