Break is available: it finds where your agent fails before your users do. Read the launch
Prove is available: it tells you whether a change to your agent changed how often it fails. Read the launch
Poisson Labs

Changelog · Tasks · 2026-11-08

Tasks 0.4.1: three fixes to export, reading a card, and the uv a run uses

0.4.1

A patch release: three fixes. Nothing about the tasks moved.

export says what verifiers’ display means. The eval command that export prints is verifiers’, and so are the words it shows. [success ] beside an episode means it ended without an error, whatever it scored (the score is the reward= beside it), and the usage line counts each conversation once, so it shows less than the run spent, because each model call sends the conversation again and is billed for it. export now says so in a second note (and in --json, as notes), and the export page says where the real figure is. The display itself is verifiers’ and is not changed.

Reading a card no longer builds it, a finished run leaves no lock file, and a run keeps the uv you have. In the guided session, “Read the card”, the word card and the run view used to read a built card by running card --force, which wrote its four files again. They now draw the run’s own card/card.md as it stands, run no verb and write nothing; a card the run has gone on from (made from what ran, and the run finished since) is still built again. run removes its run.lock when it ends, whether it finishes, fails or is stopped with Ctrl-C, and only the run that holds the lock removes it. verifiers’ bootstrap ensures the newest uv before the first episode of every run, and with pip kept out of your user site, which run does on purpose, that meant astral’s installer was fetched on every run, with a uv already installed, and replaced it. A run now uses the first uv on the harness’s PATH when it takes --script in uv sync and in uv python find, which is uv 0.6.10 or later, and fetches the installer only without one. The plan’s harness line, doctor and the README say so.

A run’s card says what the run recorded. The card of a run made with run no longer says a bare not recorded for concurrency and budget: for the concurrency it says that the figure is in the run’s run.log, and for the budget it says why there is none. It no longer carries the paragraph on resumed runs when the run was never resumed, and the card of a run on your own tools says what an intent is where it uses the word. init says that the estimate it shows before the cap, on your own tools, is the published card’s usage on the built-in tasks scaled to the size of your tools, and so does the estimate row of the plan run prints. The published cards and their page are unchanged.

Nothing about the tasks moved. Both generator hashes, and so every task id, are what 0.4.0 had. pip install -U poisson-tasks.

MATH ASINFRASTRUCTURE.

Talk to us