@yeaft/webchat-agent 0.1.120 → 0.1.122

package/claude.js

@@ -13,6 +13,22 @@ import {
   writeBackRouteContext
 } from './roleplay.js';
 
+/**
+ * Determine maxContextTokens and autoCompactThreshold from model name.
+ * Returns defaults suitable for the model's context window size.
+ */
+export function getModelContextConfig(modelName) {
+  if (!modelName) return { maxContext: 128000, compactThreshold: 110000 };
+  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 };
+}
+
 /**
  * Start a Claude SDK query for a conversation
  * Uses the SDK with AsyncIterable input stream for bidirectional communication
@@ -288,8 +304,13 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
         state.tools = message.tools || [];
         state.slashCommands = message.slash_commands || [];
         state.model = message.model || null;
+        // Set per-conversation context config based on model
+        const modelConfig = getModelContextConfig(state.model);
+        state.maxContextTokens = modelConfig.maxContext;
+        state.autoCompactThreshold = modelConfig.compactThreshold;
         console.log(`Claude session ID: ${state.claudeSessionId}`);
         console.log(`Model: ${state.model}`);
+        console.log(`Model context: ${state.maxContextTokens} tokens, compact threshold: ${state.autoCompactThreshold}`);
         console.log(`Available tools: ${state.tools.length}`);
         console.log(`Tools: ${state.tools.join(', ')}`);
         console.log(`Available slash commands: ${state.slashCommands.join(', ')}`);
@@ -420,7 +441,7 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
 
         // 计算上下文使用百分比
         const inputTokens = message.usage?.input_tokens || 0;
-        const maxContextTokens = ctx.CONFIG?.maxContextTokens || 128000;
+        const maxContextTokens = state.maxContextTokens || ctx.CONFIG?.maxContextTokens || 128000;
         if (inputTokens > 0) {
           ctx.sendToServer({
             type: 'context_usage',
@@ -527,7 +548,7 @@ async function processClaudeOutput(conversationId, claudeQuery, state) {
             });
 
             // ★ Pre-send compact check for RolePlay auto-continue
-            const rpAutoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 100000;
+            const rpAutoCompactThreshold = state.autoCompactThreshold || ctx.CONFIG?.autoCompactThreshold || 110000;
             const rpEstimatedNewTokens = Math.ceil(prompt.length / 3);
             // Include output_tokens: the assistant's output becomes part of context for the next turn
             const rpOutputTokens = message.usage?.output_tokens || 0;

package/conversation.js

@@ -673,7 +673,7 @@ export async function handleUserInput(msg) {
   console.log(`[${conversationId}] Sending: ${prompt.substring(0, 100)}...`);
 
   // ★ Pre-send compact check: estimate total tokens and compact before sending if needed
-  const autoCompactThreshold = ctx.CONFIG?.autoCompactThreshold || 100000;
+  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

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.120",
+  "version": "0.1.122",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",