Hermes Agent Token 机制详细说明

一、起点：Agent 的输出有 token 限制吗？

再看第一个问题是，Hermes Agent 有没有对 agent 输出内容做 max token 设置？答案是有，而且是四层防线协作。

从第一层来看，请求时设置 max_tokens 上限

每次调 API 时通过 _max_tokens_param 把上限传给模型。参数名因 provider 不同：

# run_agent.py:- — _max_tokens_param 定义def _max_tokens_param -> dict:
"""Return correct max tokens kwarg for current provider.
OpenAI's newer models require
'max_completion_tokens'. Azure OpenAI also requires
'max_completion_tokens' for gpt-.x models. OpenRouter。local models,and older OpenAI models use 'max_tokens'.
"""
if self._is_direct_openai_url or self._is_azure_openai_url or self._is_github_copilot_url:
return {"max_completion_tokens": value}
return {"max_tokens": value}

来源是 agent.max_tokens默认从 models.dev 自动获取模型上限。可在 config.yaml 里覆盖：

# agent_init.py:- — config.yaml 中的 max_tokens 覆盖
# Read explicit model output-token override from config when
# caller did not pass one directly.
_model_cfg = _agent_cfg.get
if agent.max_tokens is None and isinstance:
_config_max_tokens = _model_cfg.get
if _config_max_tokens is not None:
再看try，if isinstance:
raise ValueError
_parsed_max_tokens = int
if _parsed_max_tokens <= 0:
                raise ValueError
            agent.max_tokens = _parsed_max_tokens
        except :
            _ra.logger.warning(
                "Invalid model.max_tokens in config.yaml: %r — "
                "must be a positive integer . "
                "Falling back to provider default.",
                _config_max_tokens,
            )
agent._session_init_model_config = agent.max_tokens

至于第二层。检测截断

API 返回后检查 finish_reason

# conversation_loop.py:- —
if status == "incomplete" and incomplete_reason in {"max_output_token_limit_exceeded","length"}:
finish_reason = "length"
if finish_reason == "length":
# ⚠️ 输出被截断了

还额外检测 Ollama/GLM 的“假完成”——有些模型明明被截断了却返回 "stop"Hermes 用 第三层这方面，自动续写

检测到截断后自动发一条续写指令让模型从断点继续。 详见第二章，