You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Right now users blindly approach the context window limit, and then run into it, breaking their current Goose session. It would be helpful to expose the remaining context window length in some fashion. Yes this varies by LLM provider, but at least with the default provider (currently Claude) we likely have a good understanding of the limit.
With some kind of feedback about approaching the context window limit, or how much of the window got consumed up to this point (or on each message), users could better pace the session. Users could learn that certain commands consume large amounts of the context window. They could also learn how much of the context window is consumed before they're even able to ask the first question: with extensions (like Memory), and goosehints files. If a session approaches the limit, then users would know to wrap things up and/or prepare to start up a fresh session.
For the GUI, I'm envisioning some kind of bar with number-of-tokens remaining (or % remaining) that appropriately drops after each turn. For the CLI, perhaps a small text note at the end of each turn with the number-of-tokens remaining (or % remaining). If necessary, we could also have a setting to hide/show this information. Alternately, maybe a "you've consumed 75% of the context window" warning could help (with a potentially configurable threshold).
The text was updated successfully, but these errors were encountered:
Right now users blindly approach the context window limit, and then run into it, breaking their current Goose session. It would be helpful to expose the remaining context window length in some fashion. Yes this varies by LLM provider, but at least with the default provider (currently Claude) we likely have a good understanding of the limit.
With some kind of feedback about approaching the context window limit, or how much of the window got consumed up to this point (or on each message), users could better pace the session. Users could learn that certain commands consume large amounts of the context window. They could also learn how much of the context window is consumed before they're even able to ask the first question: with extensions (like Memory), and goosehints files. If a session approaches the limit, then users would know to wrap things up and/or prepare to start up a fresh session.
For the GUI, I'm envisioning some kind of bar with number-of-tokens remaining (or % remaining) that appropriately drops after each turn. For the CLI, perhaps a small text note at the end of each turn with the number-of-tokens remaining (or % remaining). If necessary, we could also have a setting to hide/show this information. Alternately, maybe a "you've consumed 75% of the context window" warning could help (with a potentially configurable threshold).
The text was updated successfully, but these errors were encountered: