跳转到内容

Deliberate Tool Mode

此内容尚不支持你的语言。

Deliberate Tool Mode is experimental. It shows the model task tools only when a message appears to need them. This keeps the prompt smaller and can help local or smaller models respond faster. The model still decides whether to use an available tool.

TomoriBot uses keywords and regular-expression patterns to match common requests, such as searching the web, setting reminders, and generating images. This is a best-effort match: an unfamiliar phrase may not make the right tool available. Recognized follow-ups such as “do that again” can reuse a recent tool request.

For example, these messages can make the corresponding tools available:

Search the web for local events. → web search
Generate an image of a red panda. → image generation
Do that again, but with a hat. → a recent tool follow-up

The patterns look for wording rather than understanding the request. For example, an image request can match a pattern like this one, which looks for an action followed by an image-related word:

\b(generate|create|make|draw)\b.*\b(image|picture|photo|pic|img)\b

She also checks for attachments and links in the triggering message or the message it replies to. Media elsewhere in the conversation does not trigger its reading tool. Triggers for all supported languages are checked regardless of your language setting.

Sticker selection can remain available for spontaneous replies, and short-term memory can update without a request. Normal permissions and restrictions still apply.

Server managers can enable the mode and add custom tool triggers under /config > Behavior > Experimental Behavior. Users can override the mode for themselves in /personal config.

If a log channel is configured, successful tool calls include the trigger that made the tool available.