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
This commit was created on GitHub.com and signed with GitHub’s verified signature.
How hard the model thinks
A model can be told how much effort to put into its replies: low, medium, high, and for the Claude models that take them xhigh and max. Click ◔ beside the model on the status bar and pick a level, or
type /effort (/effort high picks one directly). default, which is
where every model starts, says nothing and leaves it to the provider, as
before.
The level is kept for each model, and comes back when the model is
chosen again. effort = "high" under [models."name"] in the config
file sets one to start from.
It works with Anthropic, with Claude models on Amazon Bedrock, and with
OpenAI, OpenRouter and OpenAI-compatible providers such as Ollama. A
model that refuses a level is asked again without it, goes back to its
own level, and the panel says so.
Tables in replies
A table in the agent's reply is drawn as a table, with ruled columns,
where before it was shown as rows of | that wrapped mid-cell. A table
wider than the panel wraps its longest cells, and in a very narrow
panel each row becomes a short list.
Calls that are tried again
A request the provider turns away because it is busy or rate limited,
one that meets an error on the provider's side, and one whose connection
drops are made again: up to four times, waiting 2, 4, 8 and 16 seconds
or as long as the provider asks. The panel counts down what it is
waiting for, and Esc stops the wait. What waiting cannot mend is not
waited for: a wrong key, an account out of credit, a server on this
machine that is not running.
A reply cut off part of the way through is asked for again once, and
what had been shown of it is taken back. Before, a dropped connection
could leave half a reply standing as if it were the whole.
Amazon Bedrock
Claude models on Bedrock now keep the prompt between requests (prompt
caching), so the steps of a turn cost a fraction of what they did.
Claude models that think for themselves show a summary of their
thinking, as they do with Anthropic.
OpenAI's newer models
OpenAI's own models are now asked through its Responses API, which is
the only way its newer ones use tools while thinking. Before, they
refused every request the agent made ("Function tools with
reasoning_effort are not supported"). Other OpenAI-compatible providers
are asked as before; api = "responses" or api = "chat" in a
provider's table chooses where the address does not say.
OpenRouter, ready to use
openrouter is a built-in provider. Give it a key with /key openrouter (or in OPENROUTER_API_KEY) and /model lists its
models. Each model's context length is read from OpenRouter's own list,
and an effort level is sent in the form OpenRouter takes.