Repository navigation
Turnframe 0.2.1 runs on OpenAI's GPT-6 models, and its figures are now measured on gpt-6-luna.
Published at 0.2.1: turnframe-provider-openai, turnframe-tasks, turnframe-understand, turnframe-test and the turnframe facade, which asks for the four at 0.2.1. Every other crate stays at 0.2.0. No breaking change.
Fixed
- GPT-6 models run. The OpenAI adapter knew reasoning models only by the
gpt-5,o1,o3ando4names, so agpt-6model was sentmax_tokensand a temperature, and OpenAI refused every call. Everygptfromgpt-5on is now sentmax_completion_tokensand its reasoning effort, and the least effort is one each model takes:noneongpt-6-lunaandgpt-6-sol,lowongpt-6-astraandgpt-6.1-sol. The adapter's live test passes on all four. - The reply is written at the least effort. It sent none, and a model whose default effort reasons could spend the whole reply thinking and return nothing.
- Refusing to give a value («you won't get A from me») is recorded as a refusal, and saying what one would rather a record hold («I'd rather A») is a request, not small talk.
- The check takes «the end of the month» or «next Friday» as stating the day they place, and still finds a day they do not place different.
- In the trip sample, a payer given for an extra the trip already holds assigns the payer, without asking for a new extra's price.
- Builds without warnings on Rust 1.99, which deprecates
AtomicU32::fetch_update.
Measured against OpenAI's gpt-6-luna at medium: the live corpus passed 227 of 228 samples for $0.37, merged from a whole run and re-runs of the items this release's rules act on, and ten simulated conversations reached their goal ten times with no dead end and no offer refused. The gpt-5.4-mini runs of 0.2.0 stay in the benchmarks.
What changed, crate by crate, is in the changelog. Minimum supported Rust version: 1.88. Licensed under MIT or Apache-2.0.
cargo add turnframe