feat: tighten context budget - maxlen 15k, trim refactor with staged compress

ga.py

@@ -414,7 +414,7 @@ class GenericAgentHandler(BaseHandler):
                            count=count, show_linenos=show_linenos)
         if show_linenos and not result.startswith("Error:"): result = '由于设置了show_linenos，以下返回信息为：(行号|)内容 。\n' + result 
         if ' ... [TRUNCATED]' in result: result += '\n\n（某些行被截断，如需完整内容可改用 code_run 读取）'
-        maxlen = 20000 // args.get('_tool_num', 1)
+        maxlen = 15000 // args.get('_tool_num', 1)
         result = smart_format(result, max_str_len=maxlen, omit_str='\n\n[omitted long content]\n\n')
         next_prompt = self._get_anchor_prompt(skip=args.get('_index', 0) > 0)
         log_memory_access(path)
@@ -528,7 +528,7 @@ class GenericAgentHandler(BaseHandler):
                 flush(); parts.append(line); cnt = 0; last = ''
             else: cnt += 1; last = line
         flush()
-        return "\n".join(parts[-150:])
+        return "\n".join(parts[-100:])
 
     def _get_anchor_prompt(self, skip=False):
         if skip: return "\n"

llmcore.py

@@ -88,18 +88,19 @@ def safeprint(*argv):
 print = safeprint
 
 def trim_messages_history(history, context_win):
+    cap = context_win * 3
+    target = int(cap * 0.6)
+    def cost(): return sum(len(json.dumps(m, ensure_ascii=False)) for m in history)
     compress_history_tags(history)
-    cost = sum(len(json.dumps(m, ensure_ascii=False)) for m in history) 
-    print(f'[Debug] Current context: {cost} chars, {len(history)} messages.')
-    if cost > context_win * 3: 
-        compress_history_tags(history, keep_recent=4, force=True)   # trim breaks cache, so compress more btw
-        target = context_win * 3 * 0.6
-        while len(history) > 5 and cost > target:
-            history.pop(0)
-            while history and history[0].get('role') != 'user': history.pop(0)
-            if history and history[0].get('role') == 'user': history[0] = _sanitize_leading_user_msg(history[0])
-            cost = sum(len(json.dumps(m, ensure_ascii=False)) for m in history)
-        print(f'[Debug] Trimmed context, current: {cost} chars, {len(history)} messages.')
+    print(f'[Debug] Current context: {cost()} chars, {len(history)} messages.')
+    if cost() <= cap: return
+    compress_history_tags(history, keep_recent=4, force=True)
+    if cost() <= target: return
+    while len(history) > 9 and cost() > target:
+        history.pop(0)
+        while history and history[0].get('role') != 'user': history.pop(0)
+        if history and history[0].get('role') == 'user': history[0] = _sanitize_leading_user_msg(history[0])
+    print(f'[Debug] Trimmed context, current: {cost()} chars, {len(history)} messages.')
 
 def auto_make_url(base, path):
     b, p = base.rstrip('/'), path.strip('/')
@@ -510,7 +511,7 @@ class BaseSession:
         self.api_key = cfg['apikey']
         self.api_base = cfg['apibase'].rstrip('/')
         self.model = cfg.get('model', '')
-        self.context_win = cfg.get('context_win', 28000)
+        self.context_win = cfg.get('context_win', 30000)
         self.history = []
         self.lock = threading.Lock()
         self.system = ""