This commit is contained in:
fengxiangdebaozi committed 2026-05-23 17:23:22 +08:00
commit 3987ee7ee3
48 files changed
+1858 -334

No files matched your search

@@ -171,21 +171,21 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
def _build_retry_message(self, attempt: int, wait_ms: int, reason: str) -> str:
seconds = max(1, round(wait_ms / 1000))
reason_text = "provider is busy" if reason == "busy" else "provider request failed temporarily"
return f"LLM request retry {attempt}/{self.retry_max_attempts}: {reason_text}. Retrying in {seconds}s."
reason_text = "服务繁忙" if reason == "busy" else "网络/连接异常"
return f"大模型因{reason_text}调用失败,正在第 {attempt}/{self.retry_max_attempts} 次重试,{seconds} 秒后继续…"
def _build_circuit_breaker_message(self) -> str:
return "The configured LLM provider is currently unavailable due to continuous failures. Circuit breaker is engaged to protect the system. Please wait a moment before trying again."
return "大模型因连续多次调用失败已被熔断保护,请稍候再试。"
def _build_user_message(self, exc: BaseException, reason: str) -> str:
detail = _extract_error_detail(exc)
if reason == "quota":
return "The configured LLM provider rejected the request because the account is out of quota, billing is unavailable, or usage is restricted. Please fix the provider account and try again."
return "大模型因账户额度不足、账单异常或使用受限而拒绝请求,请检查模型账户后再试。"
if reason == "auth":
return "The configured LLM provider rejected the request because authentication or access is invalid. Please check the provider credentials and try again."
return "大模型因鉴权失败或无访问权限而拒绝请求,请检查模型的 API Key 与权限配置后再试。"
if reason in {"busy", "transient"}:
return "The configured LLM provider is temporarily unavailable after multiple retries. Please wait a moment and continue the conversation."
return f"LLM request failed: {detail}"
return "大模型因服务繁忙或网络异常,多次重试后仍未恢复,请稍候再继续。"
return f"大模型因「{detail}」原因调用失败。"
def _emit_retry_event(self, attempt: int, wait_ms: int, reason: str) -> None:
try:
@@ -117,15 +117,18 @@ class MemoryMiddleware(AgentMiddleware[MemoryMiddlewareState]):
except RuntimeError:
return None
def _resolve_agent_id(self, runtime: Runtime | None) -> str:
"""解析 agent_id;优先 middleware 构造时的 agent_name,再 runtime context。"""
def _resolve_agent_id(self, runtime: Runtime | None) -> str | None:
"""解析 agent_id;优先 middleware 构造时的 agent_name,再 runtime context。
无 agent 上下文时返回 None,使 V1 记忆存入全局文件(users/{id}/memory.json)。
"""
if self._agent_name:
return self._agent_name
if runtime is not None and getattr(runtime, "context", None):
agent_id = runtime.context.get("agent_id")
if agent_id:
return str(agent_id)
return "default"
return None
# ---- wrap_model_call:Hindsight 动态召回注入 -------------------------
@@ -105,8 +105,16 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
return prompt, user_msg
def _strip_think_tags(self, text: str) -> str:
"""Remove <think>...</think> blocks emitted by reasoning models (e.g. minimax, DeepSeek-R1)."""
return re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.IGNORECASE).strip()
"""Remove <think>...</think> blocks emitted by reasoning models (e.g. minimax, GLM, DeepSeek-R1).
Also handles malformed cases: an unclosed `<think>` (response truncated mid-reasoning)
or a stray `</think>` (open tag eaten by upstream parsing) — both can otherwise leak
raw reasoning into the title.
"""
text = re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
text = re.sub(r"<think>[\s\S]*$", "", text, flags=re.IGNORECASE)
text = re.sub(r"^[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
return text.strip()
def _parse_title(self, content: object) -> str:
"""Normalize model output into a clean title string."""
@@ -146,6 +154,36 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
_, user_msg = self._build_title_prompt(state)
return {"title": self._fallback_title(user_msg)}
def _force_disable_thinking(self, model: object) -> None:
"""Force-disable thinking on the title model via `extra_body` overrides.
Title generation never needs reasoning. The model factory only emits
disable params when the model config declares `when_thinking_disabled`
or a matching `when_thinking_enabled` shape — many user-added reasoning
models (e.g. GLM-4.5/5) only set `supports_thinking: true` and would
otherwise keep thinking on, which is slow and leaks <think> tags into
the title prompt response. Injecting both common OpenAI-compatible
disable shapes covers Zhipu GLM (`thinking.type`) and Qwen/vLLM
(`chat_template_kwargs.enable_thinking`); other providers ignore
unknown body fields.
"""
if not hasattr(model, "extra_body"):
return
current = getattr(model, "extra_body", None) or {}
merged_chat_template_kwargs = {
**(current.get("chat_template_kwargs") or {}),
"enable_thinking": False,
}
merged = {
**current,
"thinking": {**(current.get("thinking") or {}), "type": "disabled"},
"chat_template_kwargs": merged_chat_template_kwargs,
}
try:
model.extra_body = merged
except Exception:
logger.debug("Could not override extra_body to disable thinking on title model", exc_info=True)
async def _agenerate_title_result(self, state: TitleMiddlewareState) -> dict | None:
"""Generate a title asynchronously and fall back locally on failure."""
if not self._should_generate_title(state):
@@ -162,6 +200,7 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
model = create_chat_model(name=config.model_name, **model_kwargs)
else:
model = create_chat_model(**model_kwargs)
self._force_disable_thinking(model)
response = await model.ainvoke(prompt, config=self._get_runnable_config())
title = self._parse_title(response.content)
if title: