@yeaft/webchat-agent 0.1.161 → 0.1.163

package/claude.js

@@ -5,17 +5,22 @@ import { sendConversationList, sendOutput, sendError, handleAskUserQuestion } fr
 /**
  * Determine maxContextTokens and autoCompactThreshold from model name.
  * Returns defaults suitable for the model's context window size.
+ *
+ * NOTE (2026-03): Opus 4.6 / Sonnet 4 have 200k context windows.
+ * Claude Code handles its own compaction internally, so we set the
+ * default threshold to 200k (effectively never triggers our custom compact).
+ * The thresholds are kept as parameters in case we need to re-enable later.
  */
 export function getModelContextConfig(modelName) {
-  if (!modelName) return { maxContext: 128000, compactThreshold: 110000 };
+  if (!modelName) return { maxContext: 200000, compactThreshold: 200000 };
   const name = modelName.toLowerCase();
   // Explicit 1M context indicators
   if (name.includes('1m') || name.includes('1000k')) {
     return { maxContext: 1000000, compactThreshold: 256000 };
   }
-  // Default: 128k — Copilot API models (Sonnet 4, Opus 4, Claude 3.5 etc.)
-  // report 200k context but actual usable window is 128k
-  return { maxContext: 128000, compactThreshold: 110000 };
+  // Default: 200k — Opus 4.6 / Sonnet 4 context window.
+  // Claude Code manages its own compaction; we no longer need custom compact logic.
+  return { maxContext: 200000, compactThreshold: 200000 };
 }
 
 /**
@@ -484,39 +489,41 @@ 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}`);
-      }
+    // DISABLED (2026-03): Opus 4.6 has 200k context. Claude Code handles its own compaction.
+    // Keeping code for reference; re-enable if we ever need custom overflow recovery.
+    // } 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/conversation.js

@@ -6,6 +6,20 @@ import { crewSessions, loadCrewIndex } from './crew.js';
 // 不支持的斜杠命令（真正需要交互式 CLI 的命令）
 const UNSUPPORTED_SLASH_COMMANDS = ['/help', '/bug', '/login', '/logout', '/terminal-setup', '/vim', '/config'];
 
+/**
+ * Prestart Claude CLI process in background (fire-and-forget).
+ * When the query starts, processClaudeOutput will receive the system init message
+ * containing skills/tools/model and push them to the frontend immediately.
+ * This eliminates the delay where users had to send a message first.
+ *
+ * Errors are silently caught — failure just degrades to lazy-start behavior.
+ */
+function prestartClaude(conversationId, workDir, resumeSessionId) {
+  startClaudeQuery(conversationId, workDir, resumeSessionId).catch(err => {
+    console.warn(`[Prestart] Failed for ${conversationId}: ${err.message}`);
+  });
+}
+
 /**
  * 解析斜杠命令
  * @param {string} message - 用户消息
@@ -163,6 +177,10 @@ export async function createConversation(msg) {
   }
 
   sendConversationList();
+
+  // ★ Prestart Claude CLI in background to eagerly fetch skills/tools/model
+  // Fire-and-forget: failure just degrades to lazy-start behavior
+  prestartClaude(conversationId, effectiveWorkDir, null);
 }
 
 // Resume 历史 conversation (延迟启动 Claude，等待用户发送第一条消息)
@@ -242,6 +260,13 @@ export async function resumeConversation(msg) {
   }
 
   sendConversationList();
+
+  // ★ Prestart Claude CLI in background to eagerly fetch skills/tools/model
+  // Skip if conversation already has an active query (shouldn't happen, but safety check)
+  const resumeState = ctx.conversations.get(conversationId);
+  if (!resumeState?.query) {
+    prestartClaude(conversationId, effectiveWorkDir, claudeSessionId);
+  }
 }
 
 // 删除 conversation
@@ -443,35 +468,37 @@ export async function handleUserInput(msg) {
 
   console.log(`[${conversationId}] Sending: ${prompt.substring(0, 100)}...`);
 
+  // DISABLED (2026-03): Opus 4.6 has 200k context. Claude Code handles its own compaction.
+  // Keeping code for reference; re-enable if we ever need custom pre-send compact.
   // ★ Pre-send compact check: estimate total tokens and compact before sending if needed
-  const autoCompactThreshold = state.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
-  // 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 (input: ${lastInputTokens} + output: ${lastOutputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
-    ctx.sendToServer({
-      type: 'compact_status',
-      conversationId,
-      status: 'compacting',
-      message: `Auto-compacting before send: estimated ${estimatedTotal} tokens (threshold: ${autoCompactThreshold})`
-    });
-    // Send /compact first, then the user message will be sent after compact completes
-    // by storing it as a pending message
-    state._pendingUserMessage = userMessage;
-    state._pendingDisplayMessage = displayMessage;
-    state.turnActive = true;
-    state.turnResultReceived = false;
-    sendConversationList();
-    state.inputStream.enqueue({
-      type: 'user',
-      message: { role: 'user', content: '/compact' }
-    });
-    return;
-  }
+  // const autoCompactThreshold = state.autoCompactThreshold || ctx.CONFIG?.autoCompactThreshold || 110000;
+  // const lastInputTokens = state.lastResultInputTokens || 0;
+  // const lastOutputTokens = state.lastResultOutputTokens || 0;
+  // const estimatedNewTokens = Math.ceil(effectivePrompt.length / 3); // conservative: ~3 chars per token
+  // // Include output_tokens: the assistant's last output becomes part of context for the next turn
+  // const estimatedTotal = lastInputTokens + lastOutputTokens + estimatedNewTokens;
+  //
+  // if (estimatedTotal > autoCompactThreshold && state.inputStream) {
+  //   console.log(`[${conversationId}] Pre-send compact: estimated ${estimatedTotal} tokens (input: ${lastInputTokens} + output: ${lastOutputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
+  //   ctx.sendToServer({
+  //     type: 'compact_status',
+  //     conversationId,
+  //     status: 'compacting',
+  //     message: `Auto-compacting before send: estimated ${estimatedTotal} tokens (threshold: ${autoCompactThreshold})`
+  //   });
+  //   // Send /compact first, then the user message will be sent after compact completes
+  //   // by storing it as a pending message
+  //   state._pendingUserMessage = userMessage;
+  //   state._pendingDisplayMessage = displayMessage;
+  //   state.turnActive = true;
+  //   state.turnResultReceived = false;
+  //   sendConversationList();
+  //   state.inputStream.enqueue({
+  //     type: 'user',
+  //     message: { role: 'user', content: '/compact' }
+  //   });
+  //   return;
+  // }
 
   state.turnActive = true;
   state.turnResultReceived = false; // 重置 per-turn 去重标志

package/crew/routing.js

@@ -256,38 +256,40 @@ export async function dispatchToRole(session, roleName, content, fromSource, tas
     timestamp: Date.now()
   });
 
+  // DISABLED (2026-03): Opus 4.6 has 200k context. Claude Code handles its own compaction.
+  // Keeping code for reference; re-enable if we ever need custom crew pre-send compact.
   // ★ Pre-send compact check: estimate total tokens and clear+rebuild if needed
-  const autoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 100000;
-  const lastInputTokens = roleState.lastInputTokens || 0;
-  const estimatedNewTokens = Math.ceil((typeof content === 'string' ? content.length : 0) / 3);
-  const estimatedTotal = lastInputTokens + estimatedNewTokens;
-
-  if (lastInputTokens > 0 && estimatedTotal > autoCompactThreshold) {
-    console.log(`[Crew] Pre-send compact for ${roleName}: estimated ${estimatedTotal} tokens (last: ${lastInputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
-
-    // Save work summary before clearing (use lastTurnText since accumulatedText is cleared after result)
-    await saveRoleWorkSummary(session, roleName, roleState.lastTurnText || roleState.accumulatedText || '').catch(e =>
-      console.warn(`[Crew] Failed to save work summary for ${roleName}:`, e.message));
-
-    // Clear role session and rebuild
-    await clearRoleSessionId(session.sharedDir, roleName);
-    roleState.claudeSessionId = null;
-
-    if (roleState.abortController) roleState.abortController.abort();
-    roleState.query = null;
-    roleState.inputStream = null;
-
-    sendCrewMessage({
-      type: 'crew_role_cleared',
-      sessionId: session.id,
-      role: roleName,
-      contextPercentage: Math.round((lastInputTokens / (ctx.CONFIG?.maxContextTokens || 128000)) * 100),
-      reason: 'pre_send_compact'
-    });
-
-    // Recreate the query (fresh Claude process)
-    roleState = await createRoleQuery(session, roleName);
-  }
+  // const autoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 100000;
+  // const lastInputTokens = roleState.lastInputTokens || 0;
+  // const estimatedNewTokens = Math.ceil((typeof content === 'string' ? content.length : 0) / 3);
+  // const estimatedTotal = lastInputTokens + estimatedNewTokens;
+  //
+  // if (lastInputTokens > 0 && estimatedTotal > autoCompactThreshold) {
+  //   console.log(`[Crew] Pre-send compact for ${roleName}: estimated ${estimatedTotal} tokens (last: ${lastInputTokens} + new: ~${estimatedNewTokens}) exceeds threshold ${autoCompactThreshold}`);
+  //
+  //   // Save work summary before clearing (use lastTurnText since accumulatedText is cleared after result)
+  //   await saveRoleWorkSummary(session, roleName, roleState.lastTurnText || roleState.accumulatedText || '').catch(e =>
+  //     console.warn(`[Crew] Failed to save work summary for ${roleName}:`, e.message));
+  //
+  //   // Clear role session and rebuild
+  //   await clearRoleSessionId(session.sharedDir, roleName);
+  //   roleState.claudeSessionId = null;
+  //
+  //   if (roleState.abortController) roleState.abortController.abort();
+  //   roleState.query = null;
+  //   roleState.inputStream = null;
+  //
+  //   sendCrewMessage({
+  //     type: 'crew_role_cleared',
+  //     sessionId: session.id,
+  //     role: roleName,
+  //     contextPercentage: Math.round((lastInputTokens / (ctx.CONFIG?.maxContextTokens || 128000)) * 100),
+  //     reason: 'pre_send_compact'
+  //   });
+  //
+  //   // Recreate the query (fresh Claude process)
+  //   roleState = await createRoleQuery(session, roleName);
+  // }
 
   // P1-4: 守卫 stream.enqueue — stream 可能已被 abort 关闭
   roleState.lastDispatchContent = content;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.161",
+  "version": "0.1.163",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",