feat: 统一 max_input_tokens API,所有引擎必须返回最大输入 token 数

- openai/ollama/manual 三引擎均暴露 max_input_tokens 属性(统一契约)
- ollama 额外上报 prompt_eval_count 到 last_prompt_tokens,使本地模型也走 token 感知压缩
- should_compress 改用 max_input_tokens,系统依当前引擎上限的 70%(compress_ratio)自动触发
- 压缩始终调用当前引擎(current_engine)做总结,符合"调用当前引擎总结压缩"
- 新增 tests/test_max_input_tokens.py 覆盖契约 + 70% 触发 + 消息数兜底

Co-Authored-By: WorkBuddy <workbuddy@tencent.com>
This commit is contained in:
shuiyu67andWorkBuddy committed 2026-08-06 22:50:18 +08:00
1 parent c87a07706e
commit 5fe1a7c868
6 files changed
+104 -7

No files matched your search

+7
View File
@@ -44,6 +44,8 @@ class ManualAI:
self.conversation_history = [] self.conversation_history = []
self.max_history = 10 self.max_history = 10
self._subprocess = None self._subprocess = None
# 最近一次请求的 prompt token 数(手动引擎无真实模型,恒为 None)
self.last_prompt_tokens = None
self._session_id = str(int(time.time())) self._session_id = str(int(time.time()))
self._started = False self._started = False
@@ -240,6 +242,11 @@ while True:
time.sleep(0.3) time.sleep(0.3)
''' '''
@property
def max_input_tokens(self):
"""当前引擎支持的最大输入 token 数(统一 API)。手动引擎无真实模型,给名义上限。"""
return 128000
def generate_response(self, user_input, tool_results=None, system_prompt=None): def generate_response(self, user_input, tool_results=None, system_prompt=None):
"""生成响应 — 转发给另一个终端的人工""" """生成响应 — 转发给另一个终端的人工"""
paths = self._get_paths() paths = self._get_paths()
+19 -1
View File
@@ -113,6 +113,9 @@ Ollama AI引擎插件帮助信息
# 共享对话历史引用(由CLI设置) # 共享对话历史引用(由CLI设置)
self.shared_conversation_history = None self.shared_conversation_history = None
# 最近一次请求的 prompt token 数(供 token 感知压缩使用)
self.last_prompt_tokens = None
# 直接启用思考标记,由主程序自动处理深度思考模型的响应 # 直接启用思考标记,由主程序自动处理深度思考模型的响应
self.thinking_start_marker = "<thinking>" self.thinking_start_marker = "<thinking>"
@@ -156,7 +159,13 @@ Ollama AI引擎插件帮助信息
except Exception as e: except Exception as e:
# 如果出现异常,说明服务未运行 # 如果出现异常,说明服务未运行
return False return False
@property
def max_input_tokens(self):
"""当前引擎支持的最大输入 token 数(统一 API,供上下文压缩使用)。
本地 Ollama 由配置 max_token_k 换算的 num_ctx 决定。"""
return self.num_ctx
def generate_response(self, user_input, tool_results=None, system_prompt=None, tools=None): def generate_response(self, user_input, tool_results=None, system_prompt=None, tools=None):
""" """
生成AI响应 - 使用Ollama API 生成AI响应 - 使用Ollama API
@@ -203,6 +212,15 @@ Ollama AI引擎插件帮助信息
# ── Function Calling: 检查 tool_calls ── # ── Function Calling: 检查 tool_calls ──
message = response.get("message", {}) message = response.get("message", {})
# 上报本轮 prompt token 数(Ollama 返回 prompt_eval_count),供 token 感知压缩
try:
if isinstance(response, dict):
_pet = response.get("prompt_eval_count")
else:
_pet = getattr(response, "prompt_eval_count", None)
self.last_prompt_tokens = int(_pet) if _pet else None
except Exception:
self.last_prompt_tokens = None
tool_calls_raw = message.get("tool_calls") tool_calls_raw = message.get("tool_calls")
if tool_calls_raw: if tool_calls_raw:
# Ollama 的 tool_calls 格式: [{"function": {"name": ..., "arguments": {...}}}] # Ollama 的 tool_calls 格式: [{"function": {"name": ..., "arguments": {...}}}]
+5
View File
@@ -582,6 +582,11 @@ OpenAI 兼容格式引擎帮助信息
return win return win
return 128000 return 128000
@property
def max_input_tokens(self):
"""当前引擎支持的最大输入 token 数(统一 API,供上下文压缩使用)。"""
return self.context_window()
def list_models(self): def list_models(self):
"""列出可用模型(调用 API 的 /models 端点)""" """列出可用模型(调用 API 的 /models 端点)"""
if not self.base_url: if not self.base_url:
+66
View File
@@ -0,0 +1,66 @@
"""max_input_tokens 统一 API + token 感知压缩 回归测试
验证:
1. 所有引擎(openai / ollama / manual)都暴露 max_input_tokens 且返回正整数(API 契约)。
2. 系统依据 max_input_tokens 的 70% 自动判定压缩(should_compress)。
"""
import os
import sys
import pytest
PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if PROJECT_ROOT not in sys.path:
sys.path.insert(0, PROJECT_ROOT)
def _engine_classes():
from ai_engines.openai_engine import OpenaiAI
from ai_engines.ollama_engine import OllamaAI
from ai_engines.manual_engine import ManualAI
return {"openai": OpenaiAI, "ollama": OllamaAI, "manual": ManualAI}
def test_all_engines_expose_max_input_tokens():
"""所有引擎必须返回 max_input_tokens(统一 API 契约)"""
for name, cls in _engine_classes().items():
assert hasattr(cls, "max_input_tokens"), f"{name} 引擎缺少 max_input_tokens 属性"
def test_max_input_tokens_positive_when_instantiated():
"""实例化后 max_input_tokens 应为正整数"""
for name, cls in _engine_classes().items():
try:
inst = cls()
except Exception as e:
pytest.skip(f"{name} 引擎实例化失败(环境/网络),跳过实例检查: {e}")
val = inst.max_input_tokens
assert isinstance(val, int) and val > 0, f"{name} 的 max_input_tokens 非法: {val!r}"
def _mgr():
from xcli_core.memory import MemoryManager
return MemoryManager(project_dir=PROJECT_ROOT)
def _conv(n):
return [{"role": "user", "content": f"msg {i}"} for i in range(n)]
def test_should_compress_token_aware_triggers_at_70pct():
mgr = _mgr()
# 7000 / 10000 = 0.7 >= 0.7 → 触发
assert mgr.should_compress(_conv(5), prompt_tokens=7000, max_input_tokens=10000, ratio=0.7) is True
def test_should_compress_token_aware_below_threshold():
mgr = _mgr()
# 6000 / 10000 = 0.6 < 0.7 → 不触发(消息数也未超阈值)
assert mgr.should_compress(_conv(5), prompt_tokens=6000, max_input_tokens=10000, ratio=0.7) is False
def test_should_compress_message_count_fallback():
mgr = _mgr()
# 无 token 信息,仅消息数 > 50 → 兜底触发
assert mgr.should_compress(_conv(60)) is True
assert mgr.should_compress(_conv(10)) is False
+3 -2
View File
@@ -760,10 +760,11 @@ multi 操作支持一次修改多处:
if hasattr(self, 'memory_manager') and self.memory_manager: if hasattr(self, 'memory_manager') and self.memory_manager:
engine = self.current_engine engine = self.current_engine
prompt_tokens = getattr(engine, 'last_prompt_tokens', None) if engine else None prompt_tokens = getattr(engine, 'last_prompt_tokens', None) if engine else None
# 统一 API:所有引擎必须返回 max_input_tokens(最大输入 token 数)
ctx = None ctx = None
if engine and hasattr(engine, 'context_window'): if engine:
try: try:
ctx = engine.context_window() ctx = engine.max_input_tokens
except Exception: except Exception:
ctx = None ctx = None
ratio = get_system_config('compress_ratio', 0.7) ratio = get_system_config('compress_ratio', 0.7)
+4 -4
View File
@@ -307,18 +307,18 @@ class MemoryManager:
except (TypeError, ValueError): except (TypeError, ValueError):
pass pass
def should_compress(self, conversation: list, prompt_tokens=None, context_window=None, ratio=None) -> bool: def should_compress(self, conversation: list, prompt_tokens=None, max_input_tokens=None, ratio=None) -> bool:
""" """
判断是否需要压缩上下文。 判断是否需要压缩上下文。
优先级: 优先级:
1. token 感知: prompt_tokens/context_window >= ratio 触发(OpenAI 格式可拿到真实 token 数) 1. token 感知: prompt_tokens/max_input_tokens >= ratio 触发(引擎需返回 max_input_tokens)
2. 兜底: 消息数超过 compress_threshold 2. 兜底: 消息数超过 compress_threshold
""" """
if ratio is None: if ratio is None:
ratio = self.compress_ratio ratio = self.compress_ratio
if prompt_tokens and context_window: if prompt_tokens and max_input_tokens:
try: try:
if int(context_window) > 0 and (int(prompt_tokens) / int(context_window)) >= ratio: if int(max_input_tokens) > 0 and (int(prompt_tokens) / int(max_input_tokens)) >= ratio:
return True return True
except (TypeError, ValueError): except (TypeError, ValueError):
pass pass