Releases
1.9.9
For you
- Nothing changes for you. The private assistant tools that read these forecasts now answer with a short readable summary, such as a rankings table or a game card with its win probability and ranges exactly as published, alongside the complete data they already returned.
- A finished game in the games list now reads as a scoreline: both clubs and the final score on one row, the winner in green and the loser in red, and a Hit or Miss label with the confidence we had in the side we leaned toward. A miss shows the number we were wrong at, so a 60% call that lost says Miss 60%, never the winner's 40%. A tie shows no label, because a tie has no winner. The game's own page still uses the older wording for close calls; bringing the two into agreement is the next change.
- In light mode, the status green and red are darker, so text in them is easier to read: the winner's and loser's names and scores on a finished game, the Hit and Miss labels, a team's W and L results, and error messages. The marks that use the same colours, such as the dot beside a player ruled out and the mark beside a note that data is current, darken with them. Small text needs a contrast of 4.5 to 1. Against white the old green measured 3.35 to 1, and against the grey of the selected game in the games list the old red measured 3.7 to 1. Both now measure at least 4.7 to 1 on every light background, and the red is clearly darker than the green, so a reader who cannot tell red from green still sees the winner and the loser apart. Dark mode is unchanged.
Under the hood
- The development only styleguide page gained a token playground: every colour, spacing and radius token can be edited live in the browser and exported as a diff against the site's stylesheet. It emits nothing into the published site.
- Each of the seven assistant tools now returns two text blocks instead of one: a markdown summary first, then the full result as JSON, which used to be the only block. A program that used to read the first text block as JSON now reads the second block instead, since the first block is markdown, not JSON. A first version of this change also sent that same full result a third time, in a separate field meant for a program to read directly; a check found that the assistant client this project's own tools are read through shows that separate field and hides both text blocks whenever it is present, which hid the readable summary from the person using it. That separate field is gone now, so the summary shows again, and a program still reads the full result from the second text block exactly as before. The server also negotiates protocol version 2025-06-18 when a client asks for it, or 2025-03-26 when a client asks for that. An unsupported protocol version header on a request after initialize now gets a 400 response instead of a tool result; a request with no such header is answered as before. The summaries print published values exactly and show a missing one as absent; a value that carries a record of where it came from, such as a rankings row or a game card's win probability, is labelled with that source, though the track record's call outcomes and scoring metrics and the manifest's fields carry no such record and are shown without a label. Each summary opens with the research framing and stops at a documented number of rows with a pointer to the paging argument that reaches the rest. A formatter that fails still ships the full result; only its own readable summary is replaced with a short note. A test scans the formatter's fixed wording against the site's public voice list, and another checks that the full result of every tool is identical to what the tool returned before.
- The assistant tools' summary code now describes the result it sits beside correctly: the full result travels as JSON in the second text block, not in a separate structured field, which the tools stopped sending in the previous change. No output changes.