@yeaft/webchat-agent 0.1.90 → 0.1.92

package/claude.js

@@ -147,6 +147,16 @@ export async function startClaudeQuery(conversationId, workDir, resumeSessionId)
   return state;
 }
 
+/**
+ * Detect if an error message indicates prompt token count exceeded the model limit.
+ * Matches API errors like "prompt token count of 138392 exceeds the limit of 128000".
+ */
+export function isPromptTokenOverflow(errorMessage) {
+  if (!errorMessage) return false;
+  const msg = errorMessage.toLowerCase();
+  return msg.includes('prompt') && msg.includes('token') && (msg.includes('exceed') || msg.includes('limit'));
+}
+
 /**
  * 检测并追踪后台任务（仅 Bash 和 Agent 任务）
  * 普通工具调用（Read、Edit、Grep、Glob 等）不跟踪
@@ -508,10 +518,12 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
             // ★ Pre-send compact check for RolePlay auto-continue
             const rpAutoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 110000;
             const rpEstimatedNewTokens = Math.ceil(prompt.length / 3);
-            const rpEstimatedTotal = inputTokens + rpEstimatedNewTokens;
+            // Include output_tokens: the assistant's output becomes part of context for the next turn
+            const rpOutputTokens = message.usage?.output_tokens || 0;
+            const rpEstimatedTotal = inputTokens + rpOutputTokens + rpEstimatedNewTokens;
 
             if (rpEstimatedTotal > rpAutoCompactThreshold) {
-              console.log(`[RolePlay] Pre-send compact: estimated ${rpEstimatedTotal} tokens (last: ${inputTokens} + new: ~${rpEstimatedNewTokens}) exceeds threshold ${rpAutoCompactThreshold}`);
+              console.log(`[RolePlay] Pre-send compact: estimated ${rpEstimatedTotal} tokens (input: ${inputTokens} + output: ${rpOutputTokens} + new: ~${rpEstimatedNewTokens}) exceeds threshold ${rpAutoCompactThreshold}`);
               ctx.sendToServer({
                 type: 'compact_status',
                 conversationId,
@@ -543,6 +555,7 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
               type: 'user',
               message: { role: 'user', content: prompt }
             };
+            state._lastUserMessage = userMessage; // Save for prompt-overflow retry
             sendOutput(conversationId, userMessage);
             state.inputStream.enqueue(userMessage);
 
@@ -648,6 +661,39 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
     } else if (resultHandled) {
       // Turn 已正常完成，进程退出产生的 error 不发送给用户
       console.warn(`[SDK] Ignoring post-result error for ${conversationId}: ${error.message}`);
+    } else if (isPromptTokenOverflow(error.message) && state.claudeSessionId && !state._compactRetried) {
+      // ★ 兜底：prompt token 溢出 → 自动 compact + 重试（而非暴露 raw API error 给用户）
+      console.warn(`[SDK] Prompt token overflow for ${conversationId}, auto-compact + retry`);
+      const savedSessionId = state.claudeSessionId;
+      const savedLastMsg = state._lastUserMessage;
+
+      ctx.sendToServer({
+        type: 'compact_status',
+        conversationId,
+        status: 'compacting',
+        message: 'Context too long, auto-compacting and retrying...'
+      });
+
+      // 重启 SDK（startClaudeQuery 会先 abort 当前 state，使 finally 中 isStale=true）
+      try {
+        const newState = await startClaudeQuery(conversationId, state.workDir, savedSessionId);
+        newState._compactRetried = true; // 防止无限重试
+        newState.turnActive = true;
+        newState.turnResultReceived = false;
+
+        // 先 compact，再重试原始消息（如果有的话）
+        if (savedLastMsg) {
+          newState._pendingUserMessage = savedLastMsg;
+        }
+        newState.inputStream.enqueue({
+          type: 'user',
+          message: { role: 'user', content: '/compact' }
+        });
+        sendConversationList();
+      } catch (retryError) {
+        console.error(`[SDK] Compact-retry failed for ${conversationId}:`, retryError.message);
+        sendError(conversationId, `Context too long. Auto-compact failed: ${retryError.message}`);
+      }
     } else {
       console.error(`[SDK] Error for ${conversationId}:`, error.message);
       sendError(conversationId, error.message);

package/connection/message-router.js

@@ -12,7 +12,7 @@ import {
   createConversation, resumeConversation, deleteConversation,
   handleRefreshConversation, handleCancelExecution,
   handleUserInput, handleUpdateConversationSettings, handleAskUserAnswer,
-  sendConversationList, handleCheckCrewContext
+  sendConversationList, handleCheckCrewContext, handleCheckRolePlaySessions
 } from '../conversation.js';
 import {
   createCrewSession, handleCrewHumanInput, handleCrewControl,
@@ -70,6 +70,10 @@ export async function handleMessage(msg) {
       handleCheckCrewContext(msg);
       break;
 
+    case 'check_roleplay_sessions':
+      handleCheckRolePlaySessions(msg);
+      break;
+
     case 'resume_conversation':
       await resumeConversation(msg);
       break;

package/conversation.js

@@ -3,6 +3,7 @@ import { loadSessionHistory } from './history.js';
 import { startClaudeQuery } from './claude.js';
 import { crewSessions, loadCrewIndex } from './crew.js';
 import { rolePlaySessions, saveRolePlayIndex, removeRolePlaySession, loadRolePlayIndex, validateRolePlayConfig, initRolePlayRouteState, loadCrewContext, refreshCrewContext, initCrewContextMtimes } from './roleplay.js';
+import { loadRolePlaySessionIndex } from './roleplay-session.js';
 import { initRolePlayDir, writeSessionClaudeMd, generateSessionName, getDefaultRoles, getSessionDir } from './roleplay-dir.js';
 import { addRolePlaySession, findRolePlaySessionByConversationId, setActiveRolePlaySession } from './roleplay-session.js';
 
@@ -674,11 +675,13 @@ export async function handleUserInput(msg) {
   // ★ Pre-send compact check: estimate total tokens and compact before sending if needed
   const autoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 110000;
   const lastInputTokens = state.lastResultInputTokens || 0;
+  const lastOutputTokens = state.lastResultOutputTokens || 0;
   const estimatedNewTokens = Math.ceil(effectivePrompt.length / 3); // conservative: ~3 chars per token
-  const estimatedTotal = lastInputTokens + estimatedNewTokens;
+  // Include output_tokens: the assistant's last output becomes part of context for the next turn
+  const estimatedTotal = lastInputTokens + lastOutputTokens + estimatedNewTokens;
 
   if (estimatedTotal > autoCompactThreshold && state.inputStream) {
-    console.log(`[${conversationId}] Pre-send compact: estimated ${estimatedTotal} tokens (last: ${lastInputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
+    console.log(`[${conversationId}] Pre-send compact: estimated ${estimatedTotal} tokens (input: ${lastInputTokens} + output: ${lastOutputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
     ctx.sendToServer({
       type: 'compact_status',
       conversationId,
@@ -701,6 +704,7 @@ export async function handleUserInput(msg) {
 
   state.turnActive = true;
   state.turnResultReceived = false; // 重置 per-turn 去重标志
+  state._lastUserMessage = userMessage; // Save for prompt-overflow retry
   sendConversationList(); // 在 turnActive=true 后通知 server，确保 processing 状态正确
   sendOutput(conversationId, displayMessage);
   state.inputStream.enqueue(userMessage);
@@ -817,3 +821,40 @@ export function handleCheckCrewContext(msg) {
     featureCount: crewContext.features.length,
   });
 }
+
+/**
+ * Handle check_roleplay_sessions request — check if a directory has .roleplay/session.json
+ * and return the list of existing sessions for restore.
+ */
+export function handleCheckRolePlaySessions(msg) {
+  const { projectDir, requestId } = msg;
+  if (!projectDir) {
+    ctx.sendToServer({ type: 'roleplay_sessions_result', requestId, found: false, sessions: [] });
+    return;
+  }
+  const index = loadRolePlaySessionIndex(projectDir);
+  const activeSessions = index.sessions.filter(s => s.status === 'active');
+  if (activeSessions.length === 0) {
+    ctx.sendToServer({ type: 'roleplay_sessions_result', requestId, found: false, sessions: [] });
+    return;
+  }
+  ctx.sendToServer({
+    type: 'roleplay_sessions_result',
+    requestId,
+    found: true,
+    sessions: activeSessions.map(s => ({
+      name: s.name,
+      teamType: s.teamType,
+      language: s.language,
+      conversationId: s.conversationId,
+      roles: (s.roles || []).map(r => ({
+        name: r.name,
+        displayName: r.displayName,
+        icon: r.icon || '',
+        description: r.description || '',
+      })),
+      createdAt: s.createdAt,
+      updatedAt: s.updatedAt,
+    })),
+  });
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.90",
+  "version": "0.1.92",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",