feat: support NativeToolClient and optimize tool use format for native API
This commit is contained in:
Liang Jiaqing
@@ -56,7 +56,10 @@ class GeneraticAgent:
|
||||
except: pass
|
||||
for i, s in enumerate(llm_sessions):
|
||||
if isinstance(s, dict) and 'mixin_cfg' in s:
|
||||
try: llm_sessions[i] = ToolClient(MixinSession(llm_sessions, s['mixin_cfg']))
|
||||
try:
|
||||
mixin = MixinSession(llm_sessions, s['mixin_cfg'])
|
||||
if isinstance(mixin._sessions[0], (NativeClaudeSession, NativeOAISession)): llm_sessions[i] = NativeToolClient(mixin)
|
||||
else: llm_sessions[i] = ToolClient(mixin)
|
||||
except Exception as e: print(f'[WARN] Failed to init MixinSession with cfg {s["mixin_cfg"]}: {e}')
|
||||
self.llmclients = llm_sessions
|
||||
self.lock = threading.Lock()
|
||||
|
||||
28
llmcore.py
28
llmcore.py
@@ -62,7 +62,7 @@ def trim_messages_history(history, context_win):
|
||||
print(f'[Debug] Current context: {cost} chars, {len(history)} messages.')
|
||||
if cost > context_win * 3:
|
||||
target = context_win * 3 * 0.6
|
||||
while len(history) > 4 and cost > target:
|
||||
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])
|
||||
@@ -434,8 +434,16 @@ class BaseSession:
|
||||
self.history.append({"role": "user", "content": [{"type": "text", "text": prompt}]})
|
||||
trim_messages_history(self.history, self.context_win)
|
||||
messages = self.make_messages(self.history)
|
||||
for chunk in self.raw_ask(messages, model):
|
||||
content += chunk; yield chunk
|
||||
content_blocks = None
|
||||
gen = self.raw_ask(messages, model)
|
||||
try:
|
||||
while True: chunk = next(gen); content += chunk; yield chunk
|
||||
except StopIteration as e: content_blocks = e.value or []
|
||||
print(f"[DEBUG BaseSession.ask] content_blocks: {content_blocks}")
|
||||
for block in (content_blocks or []):
|
||||
if block.get('type', '') == 'tool_use':
|
||||
tu = {'name': block.get('name', ''), 'arguments': block.get('input', {})}
|
||||
yield f'<tool_use>{json.dumps(tu, ensure_ascii=False)}</tool_use>'
|
||||
if not content.startswith("Error:"): self.history.append({"role": "assistant", "content": [{"type": "text", "text": content}]})
|
||||
return _ask_gen() if stream else ''.join(list(_ask_gen()))
|
||||
|
||||
@@ -448,11 +456,13 @@ class ClaudeSession(BaseSession):
|
||||
headers = {"x-api-key": self.api_key, "Content-Type": "application/json", "anthropic-version": "2023-06-01", "anthropic-beta": "prompt-caching-2024-07-31"}
|
||||
payload = {"model": model, "messages": messages, "temperature": temperature, "max_tokens": max_tokens, "stream": True}
|
||||
if self.system: payload["system"] = [{"type": "text", "text": self.system, "cache_control": {"type": "persistent"}}]
|
||||
content_blocks = []
|
||||
try:
|
||||
with requests.post(auto_make_url(self.api_base, "messages"), headers=headers, json=payload, stream=True, timeout=(5,30)) as r:
|
||||
r.raise_for_status()
|
||||
yield from _parse_claude_sse(r.iter_lines())
|
||||
content_blocks = yield from _parse_claude_sse(r.iter_lines())
|
||||
except Exception as e: yield f"Error: {str(e)}"
|
||||
return content_blocks or []
|
||||
def make_messages(self, raw_list):
|
||||
msgs = [{"role": m['role'], "content": list(m['content'])} for m in raw_list]
|
||||
c = msgs[-1]["content"]
|
||||
@@ -462,10 +472,10 @@ class ClaudeSession(BaseSession):
|
||||
class LLMSession(BaseSession):
|
||||
def raw_ask(self, messages, model=None, temperature=0.5):
|
||||
if model is None: model = self.default_model
|
||||
yield from _openai_stream(self.api_base, self.api_key, messages, model, self.api_mode,
|
||||
return (yield from _openai_stream(self.api_base, self.api_key, messages, model, self.api_mode,
|
||||
temperature=temperature, reasoning_effort=self.reasoning_effort,
|
||||
max_retries=self.max_retries, connect_timeout=self.connect_timeout,
|
||||
read_timeout=self.read_timeout, proxies=self.proxies)
|
||||
read_timeout=self.read_timeout, proxies=self.proxies))
|
||||
def make_messages(self, raw_list): return _msgs_claude2oai(raw_list)
|
||||
|
||||
class NativeClaudeSession(BaseSession):
|
||||
@@ -620,11 +630,11 @@ class ToolClient:
|
||||
tools_json = json.dumps(tools, ensure_ascii=False, separators=(',', ':'))
|
||||
tool_instruction = f"""
|
||||
### 交互协议 (必须严格遵守，持续有效)
|
||||
请按照以下步骤思考并行动，标签之间需要回车换行：
|
||||
请按照以下步骤思考并行动：
|
||||
1. **思考**: 在 `<thinking>` 标签中先进行思考，分析现状和策略。
|
||||
2. **总结**: 在 `<summary>` 中输出*极为简短*的高度概括的单行（<30字）物理快照，包括上次工具调用结果产生的新信息+本次工具调用意图。此内容将进入长期工作记忆，记录关键信息，严禁输出无实际信息增量的描述。
|
||||
3. **行动**: 如需调用工具，请在回复正文之后输出一个（或多个）**<tool_use>块**，然后结束，我会稍后给你返回<tool_result>块。
|
||||
格式: ```<tool_use>\n{{"name": "工具名", "arguments": {{参数}}}}\n</tool_use>\n```
|
||||
3. **行动**: 如需调用工具，请在回复正文之后输出一个（或多个）**<tool_use>块**，然后结束。
|
||||
格式: ```<tool_use>{{"name": "工具名", "arguments": {{参数}}}}</tool_use>```
|
||||
|
||||
### 可用工具库（已挂载，持续有效）
|
||||
{tools_json}
|
||||
|
||||
@@ -8,12 +8,13 @@
|
||||
|
||||
# ── Mixin (实验性) ───────────────────────────────────────────────────────────────
|
||||
# key命名含 'mixin' 触发 MixinSession：多key/endpoint自动fallback + 指数退避重试
|
||||
# 约束：引用的session须同类型，不支持Native
|
||||
# mixin_config = {'llm_nos': [1, 2], 'max_retries': 3, 'base_delay': 1.5} # 序号含自身（此处mixin=0）
|
||||
# 约束：引用的session须同为Native或非Native
|
||||
# mixin_config = {'llm_nos': ['modela', 'xxxx'], 'max_retries': 5, 'base_delay': 1.5} # name匹配，含自身
|
||||
|
||||
# ── OpenAI-compatible (chat/completions or responses API) ──────────────────────
|
||||
# key命名含 'oai' 触发 LLMSession
|
||||
oai_config = {
|
||||
'name': 'modela', # 可选
|
||||
'apikey': 'sk-...',
|
||||
'apibase': 'http://your-proxy:2001',
|
||||
'model': 'openai/gpt-5.1',
|
||||
@@ -37,11 +38,11 @@ oai_config2 = {
|
||||
# ── Claude via OpenAI-compatible proxy ─────────────────────────────────────────
|
||||
# key命名含 'claude'（不含'native'）触发 ClaudeSession（走OpenAI兼容层）
|
||||
claude_config = {
|
||||
'name': 'xxxx', # 可选
|
||||
'apikey': 'sk-...',
|
||||
'apibase': 'http://your-proxy:2001',
|
||||
'model': 'claude-opus',
|
||||
# 'context_win': 12000,
|
||||
# 'prompt_cache': False,
|
||||
}
|
||||
|
||||
# ── Claude Native API ───────────────────────────────────────────────────────────
|
||||
@@ -50,8 +51,9 @@ claude_config = {
|
||||
native_claude_config = {
|
||||
'apikey': 'sk-ant-...', # Anthropic原生apikey
|
||||
'apibase': 'https://api.anthropic.com',
|
||||
'model': 'claude-opus-4-5',
|
||||
'model': 'claude-opus-4-6',
|
||||
# 'context_win': 24000,
|
||||
# 'no_system_prompt': True # 是否不使用系统提示而是使用用户消息，为了绕过cc MAX检测
|
||||
}
|
||||
|
||||
# ── OpenAI-compatible Native API ─────────────────────────────────────────────
|
||||
|
||||
Reference in New Issue
Block a user