You switch models, start a few new sessions, and later go looking for the one where you discussed the deploy plan. The sidebar shows a full sentence: 好的，这个可以做。不过我现在看不到代码，所以先把改造方案列出来…. That is not a title, it is the model’s answer. And every new session looks like that, stretching the list down the screen.

This is more than a chatty model. The model did ignore the instruction to return a short title, but Hermes has a guard built specifically for that failure — it just cannot count Chinese. The problem was reported on October 20, 2026 (issue #141969). Here is the mechanism, the evidence, and three workarounds that work today.

How a title gets made

Titles are generated in two steps. First comes an instant title, derived locally from your opening message without calling a model. Second comes a model upgrade: a background call gets a preview of the conversation and is asked to return a short title in a JSON field.

Step two normally behaves. But chatty models sometimes ignore the format and answer the user instead, returning a whole reply as the field’s content. That is what the code calls answer-shaped output.

The blind spot: len(title.split())

To catch that, agent/title_generator.py has a simple ceiling:

_MAX_TITLE_WORDS = 12
...
# Answer-shaped output guard: a title with many words is a model that ignored
# the task and answered the user's message instead.
if title is not None and len(title.split()) > _MAX_TITLE_WORDS:
    return None

The logic is reasonable: a genuine title does not run past twelve words, while a reply usually runs far beyond. The problem is str.split(), which splits on whitespace. English, German and French are word-delimited and behave; Chinese, Japanese and Korean are not, so a whole sentence tokenizes into one, three, maybe eight “words”.

The issue’s measurements come straight from a real state.db:

What got stored as the title	chars	len(split())	12-word guard
好的，这个可以做。不过我现在看不到 hermesbubble 的代码…	76	3	passes
可以，直接在这说。先判断一个关键点：登录页本身能打开吗？	32	1	passes
先确认一下：你说的 Hermes 是哪个项目/镜像？	80	8	passes
我是 ChatGPT，由 OpenAI 开发。当前对话里没有提供…	73	4	passes

On the same machine, older sessions created before the provider switch all had proper 17–24 character titles. In other words: in space-delimited languages the guard works; in Chinese it effectively does not exist. It is worse than useless in one respect — the accepted reply is persisted with title_source='llm', indistinguishable from a real title except that you remember what it was.

The proposed direction

The issue proposes making the guard script-aware: keep the word count for space-delimited text, and add a character ceiling for CJK runs. A title is typically 10–25 Chinese characters while a reply runs past 40, so 30 leaves room for a multi-clause title:

_MAX_TITLE_WORDS = 12
# CJK has no spaces, so len(split()) badly undercounts. A title stays short in
# characters too; a reply does not. ~30 leaves room for a real multi-clause title.
_MAX_TITLE_CJK_CHARS = 30

def _cjk_char_count(text: str) -> int:
    return sum(1 for ch in text if
               "\u4e00" <= ch <= "\u9fff"      # CJK Unified Ideographs
               or "\u3040" <= ch <= "\u30ff"   # Hiragana/Katakana
               or "\uac00" <= ch <= "\ud7a3")  # Hangul

To be clear about status: as of writing, that issue is still open and no patch has merged. So what follows are workarounds you can use now, not a release note.

Three things you can do now

One: rename the session. hermes sessions rename takes a session id and the new title:

hermes sessions rename 8f3c1d2a "deploy plan discussion"

For the history that already got a sentence as its title, this fixes them one at a time. Session ids come from hermes sessions list, and a prefix is enough.

Two: keep the instant title and stop the model upgrade. The instant title is derived from your own message, so it is short by construction. Turn off the background model call:

hermes config set auxiliary.title_generation.model_upgrade_enabled false

The trade-off is plainer titles — they reflect your first sentence rather than summarizing the conversation.

Three: disable auto-titles entirely. If you name sessions yourself anyway:

hermes config set auxiliary.title_generation.enabled false

New sessions then get no automatic title and you rename them yourself. Note that auxiliary.title_generation.language controls the title language (empty means follow the conversation) — it does not fix length, so do not expect it to route around this bug.

The general lesson

“Estimate length by counting words” is reflexive in a monolingual codebase and silently wrong in a multilingual product: no error, no warning, just a check that quietly passes for some writing systems. Any guard that estimates length or complexity by splitting on whitespace should ask what happens when the language has no spaces.

A few related areas are worth knowing if you work in Chinese or another CJK language: session boundaries and archiving are covered in the session boundaries guide; export, rename and --resume are documented in the hermes sessions reference; and if your UI language differs from the model’s, the language packs guide explains which surfaces follow the setting.

The short version

A session title became a full reply because the length guard counts words with split(), which is blind to CJK text: 76 Chinese characters count as three words, so a twelve-word ceiling never fires. In the short term, rename sessions with hermes sessions rename or disable auxiliary.title_generation.model_upgrade_enabled and keep the instant title; the real fix is a script-aware guard, which has not merged yet.