infiagent 3.2.2__tar.gz → 3.2.4__tar.gz

{infiagent-3.2.2/infiagent.egg-info → infiagent-3.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: infiagent
-Version: 3.2.2
+Version: 3.2.4
 Summary: InfiAgent multi-agent framework for long-running task automation
 Home-page: https://github.com/ChenglinPoly/Multi-Level-Agent
 Author: Chenglin Yu
@@ -55,7 +55,7 @@ Requires-Dist: psutil>=5.9.0
   <h1>MLA V3 - Build Domain-Specific SOTA-Level AI Agents</h1>
 
   <p>
-    <img src="https://img.shields.io/badge/version-3.2.2-blue.svg" alt="Version">
+    <img src="https://img.shields.io/badge/version-3.2.4-blue.svg" alt="Version">
     <img src="https://img.shields.io/badge/python-3.9+-green.svg" alt="Python">
     <img src="https://img.shields.io/badge/license-GPL-blue.svg" alt="License: GPL">
   </p>

{infiagent-3.2.2 → infiagent-3.2.4}/README.md

@@ -4,7 +4,7 @@
   <h1>MLA V3 - Build Domain-Specific SOTA-Level AI Agents</h1>
 
   <p>
-    <img src="https://img.shields.io/badge/version-3.2.2-blue.svg" alt="Version">
+    <img src="https://img.shields.io/badge/version-3.2.4-blue.svg" alt="Version">
     <img src="https://img.shields.io/badge/python-3.9+-green.svg" alt="Python">
     <img src="https://img.shields.io/badge/license-GPL-blue.svg" alt="License: GPL">
   </p>

{infiagent-3.2.2 → infiagent-3.2.4}/core/event_handlers.py

@@ -122,6 +122,18 @@ class ConsoleLogHandler:
     def _print_run_llm_end(self, event: LlmCallEndEvent):
         safe_print(f"📥 LLM输出: {event.llm_output[:100]}...")
         safe_print(f"🔧 工具调用数量: {len(event.tool_calls)}")
+        if event.usage:
+            safe_print(
+                f"🧮 本轮tokens: prompt={event.usage.get('prompt_tokens', 0)}, "
+                f"completion={event.usage.get('completion_tokens', 0)}, "
+                f"total={event.usage.get('total_tokens', 0)}"
+            )
+        if event.cumulative_usage:
+            safe_print(
+                f"📈 累计tokens: prompt={event.cumulative_usage.get('prompt_tokens', 0)}, "
+                f"completion={event.cumulative_usage.get('completion_tokens', 0)}, "
+                f"total={event.cumulative_usage.get('total_tokens', 0)}"
+            )
         
     def _print_run_tool_start(self, event: ToolCallStartEvent):
         safe_print(f"\n🔧 执行工具: {event.tool_name}")
@@ -141,6 +153,18 @@ class ConsoleLogHandler:
 
     def _print_run_thinking_end(self, event: ThinkingEndEvent):
         safe_print(f"[{event.agent_name}] Thinking分析已更新: {event.result}")
+        if event.usage:
+            safe_print(
+                f"🧮 Thinking本轮tokens: prompt={event.usage.get('prompt_tokens', 0)}, "
+                f"completion={event.usage.get('completion_tokens', 0)}, "
+                f"total={event.usage.get('total_tokens', 0)}"
+            )
+        if event.cumulative_usage:
+            safe_print(
+                f"📈 累计tokens: prompt={event.cumulative_usage.get('prompt_tokens', 0)}, "
+                f"completion={event.cumulative_usage.get('completion_tokens', 0)}, "
+                f"total={event.cumulative_usage.get('total_tokens', 0)}"
+            )
         
     def _print_run_thinking_fail(self, event: ThinkingFailEvent):
         safe_print(f"⚠️ Thinking触发失败: {event.error_message}")
@@ -230,13 +254,25 @@ class JsonlStreamHandler:
             "error": err[:1000] if err else ""
         })
 
+    def _stream_run_llm_end(self, event: LlmCallEndEvent):
+        self.jsonl_emitter.emit({
+            "type": "llm_end",
+            "model": event.model,
+            "finish_reason": event.finish_reason,
+            "usage": event.usage or {},
+            "cumulative_usage": event.cumulative_usage or {},
+            "request_budget": event.request_budget or {},
+        })
+
     def _stream_run_thinking_end(self, event: ThinkingEndEvent):
         # 发送完整 thinking 内容（thinking agent 独立 LLM 调用，token 不经过主 agent 事件流）
         self.jsonl_emitter.emit({
             "type": "thinking_end",
             "agent": event.agent_name,
             "is_initial": event.is_initial,
-            "result": event.result
+            "result": event.result,
+            "usage": event.usage or {},
+            "cumulative_usage": event.cumulative_usage or {},
         })
     
     def _stream_run_thinking_fail(self, event: ThinkingFailEvent):

{infiagent-3.2.2 → infiagent-3.2.4}/core/events.py

@@ -72,6 +72,8 @@ class ThinkingEndEvent(AgentEvent):
     raw_output: str = ""
     raw_reasoning_content: str = ""
     finish_reason: str = ""
+    usage: Dict[str, Any] = field(default_factory=dict)
+    cumulative_usage: Dict[str, Any] = field(default_factory=dict)
     # Default arguments last
     is_forced: bool = False 
     timestamp: float = field(default_factory=time.time)
@@ -107,6 +109,9 @@ class LlmCallEndEvent(AgentEvent):
     model: str = ""
     reasoning_content: str = ""
     finish_reason: str = ""
+    usage: Dict[str, Any] = field(default_factory=dict)
+    cumulative_usage: Dict[str, Any] = field(default_factory=dict)
+    request_budget: Dict[str, Any] = field(default_factory=dict)
     # Default arguments last
     timestamp: float = field(default_factory=time.time)
 

{infiagent-3.2.2 → infiagent-3.2.4}/core/tool_executor.py

@@ -10,6 +10,7 @@ import time
 import uuid
 import asyncio
 import threading
+import os
 from typing import Dict, Any
 
 from tool_server_lite.registry import get_runtime_registry, get_runtime_registry_failures
@@ -57,6 +58,7 @@ class ToolExecutor:
         self.task_permissions = {}  # {task_id: {"auto_mode": True/False}}
         self.agent_id = ""
         self.agent_name = ""
+        self.tool_runtime_defaults = self._load_tool_runtime_defaults()
 
     def set_agent_context(self, *, agent_id: str = "", agent_name: str = ""):
         self.agent_id = str(agent_id or "")
@@ -69,6 +71,31 @@ class ToolExecutor:
             return int(((hierarchy.get(self.agent_id) or {}).get("level")) or 0)
         except Exception:
             return 0
+
+    def _load_tool_runtime_defaults(self) -> Dict[str, Dict[str, Any]]:
+        raw = os.environ.get("MLA_TOOL_RUNTIME_DEFAULTS_JSON", "").strip()
+        if not raw:
+            return {}
+        try:
+            parsed = json.loads(raw)
+            if isinstance(parsed, dict):
+                result: Dict[str, Dict[str, Any]] = {}
+                for tool_name, defaults in parsed.items():
+                    if isinstance(defaults, dict):
+                        result[str(tool_name)] = dict(defaults)
+                return result
+        except Exception as exc:
+            safe_print(f"⚠️  解析 MLA_TOOL_RUNTIME_DEFAULTS_JSON 失败: {exc}")
+        return {}
+
+    def _merge_runtime_tool_defaults(self, tool_name: str, arguments: Dict[str, Any]) -> Dict[str, Any]:
+        defaults = self.tool_runtime_defaults.get(str(tool_name), {})
+        if not defaults:
+            return arguments
+        merged = dict(arguments or {})
+        for key, value in defaults.items():
+            merged[str(key)] = value
+        return merged
     
     def _init_tools_registry(self):
         """
@@ -250,12 +277,13 @@ class ToolExecutor:
         Returns:
             执行结果字典
         """
+        merged_arguments = self._merge_runtime_tool_defaults(tool_name, arguments)
         try:
             trigger_tool_hooks(
                 when="before",
                 tool_name=tool_name,
                 task_id=task_id,
-                arguments=arguments,
+                arguments=merged_arguments,
                 agent_id=self.agent_id,
                 agent_name=self.agent_name,
                 agent_level=self._agent_level(),
@@ -271,13 +299,13 @@ class ToolExecutor:
             # 特殊处理final_output
             if tool_name == "final_output":
                 result = {
-                    "status": arguments.get("status", "success"),
-                    "output": arguments.get("output", ""),
-                    "error_information": arguments.get("error_information", "")
+                    "status": merged_arguments.get("status", "success"),
+                    "output": merged_arguments.get("output", ""),
+                    "error_information": merged_arguments.get("error_information", "")
                 }
             elif tool_type == "tool_call_agent":
                 if tool_name in self.DANGEROUS_TOOLS and not self.is_auto_mode(task_id):
-                    approved = self._request_tool_confirmation(tool_name, arguments, task_id)
+                    approved = self._request_tool_confirmation(tool_name, merged_arguments, task_id)
                     if not approved:
                         result = {
                             "status": "error",
@@ -285,15 +313,15 @@ class ToolExecutor:
                             "error_information": f"工具执行被用户拒绝: {tool_name}"
                         }
                     elif tool_config.get("_mcp"):
-                        result = call_mcp_tool(tool_config, arguments)
+                        result = call_mcp_tool(tool_config, merged_arguments)
                     else:
-                        result = self._call_direct(tool_name, arguments, task_id)
+                        result = self._call_direct(tool_name, merged_arguments, task_id)
                 elif tool_config.get("_mcp"):
-                    result = call_mcp_tool(tool_config, arguments)
+                    result = call_mcp_tool(tool_config, merged_arguments)
                 else:
-                    result = self._call_direct(tool_name, arguments, task_id)
+                    result = self._call_direct(tool_name, merged_arguments, task_id)
             elif tool_type == "llm_call_agent":
-                result = self._execute_sub_agent(tool_name, tool_config, arguments, task_id)
+                result = self._execute_sub_agent(tool_name, tool_config, merged_arguments, task_id)
             else:
                 result = {
                     "status": "error",
@@ -312,7 +340,7 @@ class ToolExecutor:
                 when="after",
                 tool_name=tool_name,
                 task_id=task_id,
-                arguments=arguments,
+                arguments=merged_arguments,
                 result=result,
                 agent_id=self.agent_id,
                 agent_name=self.agent_name,

{infiagent-3.2.2 → infiagent-3.2.4}/infiagent/sdk.py

@@ -110,6 +110,7 @@ class InfiAgent:
         fresh_enabled: Optional[bool] = None,
         fresh_interval_sec: Optional[int] = None,
         mcp_servers: Optional[List[Dict[str, Any]]] = None,
+        tool_runtime_defaults: Optional[Dict[str, Dict[str, Any]]] = None,
         tool_hooks: Optional[List[Dict[str, Any]]] = None,
         context_hooks: Optional[List[Dict[str, Any]]] = None,
         visible_skills: Optional[List[str]] = None,
@@ -173,6 +174,8 @@ class InfiAgent:
             self.runtime_env_overrides["MLA_FRESH_INTERVAL_SEC"] = str(max(0, int(fresh_interval_sec)))
         if mcp_servers is not None:
             self.runtime_env_overrides["MLA_MCP_CONFIG_JSON"] = json.dumps({"servers": mcp_servers}, ensure_ascii=False)
+        if tool_runtime_defaults is not None:
+            self.runtime_env_overrides["MLA_TOOL_RUNTIME_DEFAULTS_JSON"] = json.dumps(tool_runtime_defaults, ensure_ascii=False)
         if tool_hooks is not None:
             self.runtime_env_overrides["MLA_TOOL_HOOKS_JSON"] = json.dumps(tool_hooks, ensure_ascii=False)
         if context_hooks is not None:
@@ -199,6 +202,7 @@ class InfiAgent:
         self.fresh_enabled = fresh_enabled if fresh_enabled is not None else None
         self.fresh_interval_sec = max(0, int(fresh_interval_sec)) if fresh_interval_sec is not None else None
         self.mcp_servers = mcp_servers
+        self.tool_runtime_defaults = tool_runtime_defaults
         self.tool_hooks = tool_hooks
         self.context_hooks = context_hooks
         self.visible_skills = [str(item).strip() for item in visible_skills if str(item).strip()] if visible_skills is not None else None
@@ -258,6 +262,8 @@ class InfiAgent:
             config["fresh_interval_sec"] = self.fresh_interval_sec
         if self.mcp_servers is not None:
             config["mcp_servers"] = self.mcp_servers
+        if self.tool_runtime_defaults is not None:
+            config["tool_runtime_defaults"] = self.tool_runtime_defaults
         if self.tool_hooks is not None:
             config["tool_hooks"] = self.tool_hooks
         if self.context_hooks is not None:
@@ -320,8 +326,10 @@ class InfiAgent:
         user_history_compress_threshold_tokens: Optional[int] = None,
         user_history_recent_items: Optional[int] = None,
         max_turns: Optional[int] = None,
+        tool_runtime_defaults: Optional[Dict[str, Dict[str, Any]]] = None,
     ) -> Dict[str, Any]:
         target_task_id = self._resolve_task_id(task_id=task_id, workspace=workspace)
+        Path(target_task_id).mkdir(parents=True, exist_ok=True)
         target_agent_system = agent_system or self.default_agent_system
         target_agent_name = agent_name or self.default_agent_name
 
@@ -347,6 +355,8 @@ class InfiAgent:
             )
         if max_turns is not None:
             extra_runtime_overrides["MLA_MAX_TURNS"] = str(max(1, int(max_turns)))
+        if tool_runtime_defaults is not None:
+            extra_runtime_overrides["MLA_TOOL_RUNTIME_DEFAULTS_JSON"] = json.dumps(tool_runtime_defaults, ensure_ascii=False)
 
         with self._runtime_scope(extra_runtime_overrides or None):
             extra_event_handlers, collected_events = self._build_sdk_event_handler(

{infiagent-3.2.2 → infiagent-3.2.4/infiagent.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: infiagent
-Version: 3.2.2
+Version: 3.2.4
 Summary: InfiAgent multi-agent framework for long-running task automation
 Home-page: https://github.com/ChenglinPoly/Multi-Level-Agent
 Author: Chenglin Yu
@@ -55,7 +55,7 @@ Requires-Dist: psutil>=5.9.0
   <h1>MLA V3 - Build Domain-Specific SOTA-Level AI Agents</h1>
 
   <p>
-    <img src="https://img.shields.io/badge/version-3.2.2-blue.svg" alt="Version">
+    <img src="https://img.shields.io/badge/version-3.2.4-blue.svg" alt="Version">
     <img src="https://img.shields.io/badge/python-3.9+-green.svg" alt="Python">
     <img src="https://img.shields.io/badge/license-GPL-blue.svg" alt="License: GPL">
   </p>

{infiagent-3.2.2 → infiagent-3.2.4}/infiagent.egg-info/SOURCES.txt

@@ -387,6 +387,7 @@ tests/test_sdk_observability.py
 tests/test_sdk_runtime_paths.py
 tests/test_task_history_index.py
 tests/test_thinking_prompt_contract.py
+tests/test_token_budgeting.py
 tests/test_tool_hooks.py
 tests/test_web_ui_config_api.py
 tests/test_windows_compat.py
@@ -425,6 +426,7 @@ utils/runtime_control.py
 utils/skill_loader.py
 utils/task_history_index.py
 utils/task_runtime.py
+utils/token_budget.py
 utils/tool_hooks.py
 utils/user_paths.py
 utils/windows_compat.py

{infiagent-3.2.2 → infiagent-3.2.4}/services/action_compressor.py

@@ -3,7 +3,11 @@ from utils.windows_compat import safe_print
 # -*- coding: utf-8 -*-
 """
 历史动作压缩服务
-策略：总结历史XML + 保留最新action + 压缩最新action的大字段
+策略：总结历史XML + 保留最新action + 压缩最新action的大字段。
+
+注意：
+- 预算统计必须覆盖 assistant_content / reasoning_content，
+  否则 ReAct / thinking 文本会逃逸上下文预算。
 """
 
 import json
@@ -61,6 +65,16 @@ class ActionCompressor:
             chinese_chars = sum(1 for c in text if '\u4e00' <= c <= '\u9fff')
             other_chars = len(text) - chinese_chars
             return int(chinese_chars / 1.5 + other_chars / 4)
+
+    def count_action_context_tokens(
+        self,
+        action_history: List[Dict],
+        *,
+        thinking: str = "",
+        task_input: str = "",
+    ) -> int:
+        serialized = self._actions_to_xml(action_history)
+        return self.count_tokens(serialized + str(thinking or "") + str(task_input or ""))
     
     def compress_if_needed(
         self,
@@ -68,7 +82,8 @@ class ActionCompressor:
         max_context_window: int,
         thinking: str = "",
         task_input: str = "",
-        save_callback=None  # 添加保存回调，确保压缩后立即保存
+        save_callback=None,  # 添加保存回调，确保压缩后立即保存
+        max_action_tokens: Optional[int] = None,
     ) -> List[Dict]:
         """
         检查并压缩历史动作
@@ -93,39 +108,59 @@ class ActionCompressor:
         # 如果只有一条
         if len(action_history) == 1:
             # 检查是否需要压缩字段
-            return [self._compress_action_fields(action_history[0], max_context_window // 2)]
+            single_budget = max(256, int(max_action_tokens or max_context_window // 2))
+            return [self._compress_action_fields(
+                action_history[0],
+                single_budget,
+                thinking=thinking,
+                task_input=task_input,
+                max_context_window=max_context_window,
+            )]
         
         # 分离最新和历史
         recent_action = action_history[-1]
         historical_actions = action_history[:-1]
         
         # 计算整体token数
-        total_text = self._actions_to_xml(action_history)
-        total_tokens = self.count_tokens(total_text+thinking+task_input)
-        
+        total_tokens = self.count_action_context_tokens(
+            action_history,
+            thinking=thinking,
+            task_input=task_input,
+        )
+
+        effective_budget = int(max_action_tokens) if max_action_tokens is not None else max(
+            1024,
+            max_context_window
+            - max(1024, min(4096, int(max_context_window * 0.15)))
+            - max(512, min(4096, int(max_context_window * 0.05))),
+        )
+
         # 如果不超限，不压缩
-        if total_tokens <= max_context_window - 20000:
+        if total_tokens <= effective_budget:
             return action_history
         
-        safe_print(f"🔄 历史动作需要压缩: {total_tokens} tokens > {max_context_window - 20000}")
+        safe_print(f"🔄 历史动作需要压缩: {total_tokens} tokens > {effective_budget}")
         
         # 压缩策略：
         # 1. 历史 → 基于 thinking 和 task_input 智能总结为5k tokens
         # 2. 最新 → 压缩为max_window的50%
         
+        summary_target = max(256, min(5000, int(effective_budget * 0.35)))
+        recent_target = max(256, max(0, effective_budget - summary_target))
+
         summary_action = self._summarize_historical_xml(
             self._actions_to_xml(historical_actions),
-            target_tokens=5000,  # 历史总结固定5k tokens
+            target_tokens=summary_target,
             thinking=thinking,
             task_input=task_input,
             max_context_window=max_context_window,
             actions=historical_actions  # 传递原始 actions（用于提取图片）
         )
         
-        # 压缩最新action的大字段（50% of max_window）
+        # 压缩最新action的大字段（优先保留最近动作）
         compressed_recent = self._compress_action_fields(
             recent_action,
-            int(max_context_window * 0.5),  # 80000 * 0.5 = 40000 tokens
+            recent_target,
             thinking=thinking,
             task_input=task_input,
             max_context_window=max_context_window
@@ -147,8 +182,17 @@ class ActionCompressor:
             tool_name = action.get("tool_name", "")
             arguments = action.get("arguments", {})
             result = action.get("result", {})
+            assistant_content = str(action.get("assistant_content", "") or "")
+            reasoning_content = str(action.get("reasoning_content", "") or "")
             
             action_xml = f"<action>\n  <tool_name>{tool_name}</tool_name>\n"
+
+            if assistant_content:
+                escaped_content = assistant_content.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
+                action_xml += f"  <assistant_content>{escaped_content}</assistant_content>\n"
+            if reasoning_content:
+                escaped_reasoning = reasoning_content.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
+                action_xml += f"  <reasoning_content>{escaped_reasoning}</reasoning_content>\n"
             
             # 参数（跳过内部字段）
             for k, v in arguments.items():
@@ -541,6 +585,23 @@ class ActionCompressor:
                 compressed_action["result"]["output"] = compressed_output
                 compressed_action["result"]["_compressed"] = True
                 compressed_action["result"]["_original_tokens"] = output_tokens
+
+        for field_name in ("assistant_content", "reasoning_content"):
+            field_value = str(compressed_action.get(field_name, "") or "")
+            if not field_value:
+                continue
+            field_tokens = self.count_tokens(field_value)
+            if field_tokens > max_field_tokens:
+                safe_print(f"   🤖 LLM压缩{field_name}: {field_tokens} tokens → {max_field_tokens} tokens")
+                compressed_action[field_name] = self._llm_compress_field(
+                    field_value,
+                    max_field_tokens,
+                    action.get("tool_name", "unknown"),
+                    thinking=thinking,
+                    task_input=task_input,
+                    field_context=f"动作中的 {field_name}",
+                    max_context_window=max_context_window
+                )
         
         return compressed_action